@sema-agent/core 7.67.3 → 7.68.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +1516 -0
- package/README.md +18 -9
- package/README.zh-CN.md +16 -8
- package/dist/agents/agent-definition.d.ts +1 -1
- package/dist/agents/agent-definition.js +4 -1
- package/dist/agents/cascade.d.ts +2 -2
- package/dist/agents/cascade.js +3 -1
- package/dist/agents/cumulative-stats.d.ts +2 -1
- package/dist/agents/cumulative-stats.js +4 -1
- package/dist/agents/list-agents-tool.d.ts +30 -0
- package/dist/agents/list-agents-tool.js +45 -3
- package/dist/agents/peer-admission.d.ts +2 -2
- package/dist/agents/peer-session-drain.d.ts +2 -0
- package/dist/agents/peer-session-drain.js +1 -0
- package/dist/agents/repair-loop.js +3 -1
- package/dist/agents/retain-ledger.d.ts +61 -8
- package/dist/agents/retain-ledger.js +25 -3
- package/dist/agents/roster-store.d.ts +4 -2
- package/dist/agents/send-message-tool.d.ts +1 -1
- package/dist/agents/send-message-tool.js +20 -15
- package/dist/agents/subagent-steps.js +2 -10
- package/dist/agents/subagent.d.ts +3 -0
- package/dist/agents/subagent.js +106 -76
- package/dist/agents/tool-filter.d.ts +20 -11
- package/dist/agents/tool-filter.js +8 -4
- package/dist/agents/verify.d.ts +3 -1
- package/dist/agents/verify.js +2 -0
- package/dist/brain/anthropic.js +4 -4
- package/dist/brain/open-responses.js +3 -1
- package/dist/brain/openai.js +3 -1
- package/dist/brain/reasoning-request.d.ts +73 -19
- package/dist/brain/reasoning-request.js +164 -72
- package/dist/brain/reasoning-stamp.d.ts +26 -0
- package/dist/brain/reasoning-stamp.js +10 -0
- package/dist/brain/reasoning.d.ts +95 -162
- package/dist/brain/reasoning.js +31 -86
- package/dist/brain/route-adjudicator.d.ts +9 -7
- package/dist/brain/route-adjudicator.js +12 -10
- package/dist/brain/status-sink.d.ts +19 -64
- package/dist/brain/status-sink.js +5 -21
- package/dist/brain/stream-engine.d.ts +11 -2
- package/dist/brain/stream-engine.js +10 -0
- package/dist/core/a2a.d.ts +17 -5
- package/dist/core/a2a.js +11 -3
- package/dist/core/allow-entry.d.ts +35 -15
- package/dist/core/allow-entry.js +47 -60
- package/dist/core/ask-headline.d.ts +68 -0
- package/dist/core/ask-headline.js +23 -0
- package/dist/core/ask-origin.d.ts +36 -8
- package/dist/core/ask-origin.js +17 -10
- package/dist/core/ask-question.d.ts +34 -7
- package/dist/core/ask-question.js +10 -1
- package/dist/core/ask-unresolvable-notice.d.ts +7 -0
- package/dist/core/ask-unresolvable-notice.js +9 -3
- package/dist/core/auto-strip.d.ts +53 -0
- package/dist/core/auto-strip.js +76 -0
- package/dist/core/background-agent-store.d.ts +45 -5
- package/dist/core/background-agent-store.js +69 -15
- package/dist/core/background-shell.d.ts +2 -2
- package/dist/core/caller-text.js +2 -2
- package/dist/core/canonical-json.d.ts +13 -0
- package/dist/core/canonical-json.js +7 -0
- package/dist/core/checkpoint-features.d.ts +66 -20
- package/dist/core/checkpoint-features.js +12 -1
- package/dist/core/checkpoint-generation.d.ts +1 -1
- package/dist/core/checkpoint-queries.d.ts +1 -1
- package/dist/core/checkpoint-queries.js +4 -1
- package/dist/core/checkpoint-requirements.d.ts +25 -1
- package/dist/core/checkpoint-requirements.js +5 -1
- package/dist/core/checkpoint-row-verdict.d.ts +1 -0
- package/dist/core/checkpoint-row-verdict.js +4 -0
- package/dist/core/checkpoint-store.d.ts +66 -38
- package/dist/core/checkpoint-store.js +15 -6
- package/dist/core/child-clamp-seats.d.ts +6 -0
- package/dist/core/child-clamp-seats.js +1 -0
- package/dist/core/config-refusal-codes.d.ts +1 -1
- package/dist/core/config-refusal-codes.js +1 -0
- package/dist/core/delegation-frames.d.ts +9 -12
- package/dist/core/delegation-frames.js +9 -8
- package/dist/core/engine-capabilities.d.ts +21 -8
- package/dist/core/engine-capabilities.js +1 -1
- package/dist/core/engine-notice.d.ts +147 -34
- package/dist/core/engine-notice.js +27 -16
- package/dist/core/engine-storage-backend.d.ts +2 -1
- package/dist/core/env-budget.js +3 -3
- package/dist/core/env-knob-announce.d.ts +10 -10
- package/dist/core/env-knob-announce.js +3 -6
- package/dist/core/file-history-store.d.ts +29 -4
- package/dist/core/file-history-store.js +13 -0
- package/dist/core/frame-settlement.d.ts +2 -1
- package/dist/core/gate-fold.js +1 -1
- package/dist/core/gate-lanes.js +56 -49
- package/dist/core/governance-codes.d.ts +39 -3
- package/dist/core/governance-codes.js +61 -3
- package/dist/core/hooks.d.ts +29 -11
- package/dist/core/hooks.js +5 -2
- package/dist/core/human-input-projection.js +3 -3
- package/dist/core/input-author.d.ts +17 -3
- package/dist/core/input-author.js +15 -3
- package/dist/core/leg-identity.d.ts +10 -28
- package/dist/core/leg-identity.js +1 -21
- package/dist/core/lsp.d.ts +18 -13
- package/dist/core/lsp.js +30 -22
- package/dist/core/mailbox-store.d.ts +1 -1
- package/dist/core/mailbox-store.js +4 -4
- package/dist/core/mcp-connection-pool.d.ts +93 -4
- package/dist/core/mcp-connection-pool.js +104 -1
- package/dist/core/mcp-failure.d.ts +38 -12
- package/dist/core/mcp-failure.js +9 -13
- package/dist/core/mcp-probe.d.ts +4 -3
- package/dist/core/mcp-server-link.d.ts +4 -2
- package/dist/core/mcp-server-link.js +4 -5
- package/dist/core/mcp-server-spec.d.ts +3 -3
- package/dist/core/mcp-server-unavailable-notice.d.ts +22 -8
- package/dist/core/mcp-server-unavailable-notice.js +18 -4
- package/dist/core/mcp.d.ts +40 -6
- package/dist/core/mcp.js +39 -13
- package/dist/core/memory-engine/content-origin.js +2 -2
- package/dist/core/memory-engine/control-objects.d.ts +2 -4
- package/dist/core/memory-engine/control-objects.js +2 -10
- package/dist/core/memory-engine/delegation-settlement.js +6 -1
- package/dist/core/memory-engine/dual-root.js +3 -0
- package/dist/core/memory-engine/engine.d.ts +6 -4
- package/dist/core/memory-engine/engine.js +32 -29
- package/dist/core/memory-engine/file-backend.d.ts +20 -5
- package/dist/core/memory-engine/file-backend.js +21 -31
- package/dist/core/memory-engine/frontmatter.d.ts +2 -0
- package/dist/core/memory-engine/frontmatter.js +4 -0
- package/dist/core/memory-engine/layout.d.ts +114 -20
- package/dist/core/memory-engine/layout.js +340 -22
- package/dist/core/memory-engine/memory-deployment.d.ts +7 -5
- package/dist/core/memory-engine/mount-failed-notice.d.ts +19 -7
- package/dist/core/memory-engine/mount-failed-notice.js +43 -10
- package/dist/core/memory-engine/sync-client.js +4 -5
- package/dist/core/model-fallbacks.d.ts +5 -1
- package/dist/core/model-fallbacks.js +3 -3
- package/dist/core/model-receipt.d.ts +2 -1
- package/dist/core/model-receipt.js +12 -1
- package/dist/core/mods/kernel/dispatch.d.ts +2 -2
- package/dist/core/mods/kernel/dispatch.js +14 -6
- package/dist/core/mods/kernel/events.d.ts +8 -5
- package/dist/core/mods/kernel/events.js +2 -2
- package/dist/core/mods/kernel/index.d.ts +1 -1
- package/dist/core/mods/kernel/registration.d.ts +9 -0
- package/dist/core/mods/package.js +1 -1
- package/dist/core/mods/session-append.d.ts +4 -2
- package/dist/core/mods/session-append.js +3 -0
- package/dist/core/mods/stations.d.ts +7 -11
- package/dist/core/mods/stations.js +4 -7
- package/dist/core/mods/tool-call.d.ts +10 -8
- package/dist/core/mods/tool-call.js +15 -18
- package/dist/core/mounted-mcp.d.ts +0 -4
- package/dist/core/mounted-mcp.js +1 -1
- package/dist/core/park-noun.d.ts +24 -0
- package/dist/core/park-noun.js +21 -0
- package/dist/core/park-resumability.d.ts +60 -0
- package/dist/core/park-resumability.js +30 -0
- package/dist/core/permission-rule-provider.d.ts +24 -2
- package/dist/core/permission-rule-provider.js +11 -8
- package/dist/core/permission-rule-store.d.ts +8 -1
- package/dist/core/permission-rule-store.js +2 -3
- package/dist/core/permission-rules.d.ts +2 -1
- package/dist/core/permission-rules.js +8 -4
- package/dist/core/persisted-rule-arms.d.ts +3 -0
- package/dist/core/persisted-rule-arms.js +2 -1
- package/dist/core/present-plan-tool.d.ts +47 -3
- package/dist/core/present-plan-tool.js +62 -4
- package/dist/core/principal-name.d.ts +1 -1
- package/dist/core/principal-name.js +5 -5
- package/dist/core/process-identity.js +10 -5
- package/dist/core/protocol-http-auth.d.ts +20 -14
- package/dist/core/protocol-http-auth.js +70 -50
- package/dist/core/read-only-shell.js +9 -5
- package/dist/core/redaction-marks.d.ts +4 -26
- package/dist/core/redaction-marks.js +9 -21
- package/dist/core/remote-env.d.ts +43 -16
- package/dist/core/remote-env.js +1 -1
- package/dist/core/request-copy-projection.d.ts +1 -1
- package/dist/core/request-copy-projection.js +40 -50
- package/dist/core/resume-config-judge.js +1 -1
- package/dist/core/resume-startup-cause.d.ts +35 -18
- package/dist/core/resume-startup-cause.js +13 -10
- package/dist/core/retention.d.ts +167 -25
- package/dist/core/retention.js +100 -20
- package/dist/core/roles.d.ts +6 -0
- package/dist/core/roles.js +11 -5
- package/dist/core/rule-lane-reads.d.ts +3 -0
- package/dist/core/rule-lane-reads.js +5 -1
- package/dist/core/rule-store-fault.d.ts +70 -0
- package/dist/core/rule-store-fault.js +102 -0
- package/dist/core/runner/advertised-writable-dirs.js +16 -31
- package/dist/core/runner/announce-once-ledger.d.ts +19 -2
- package/dist/core/runner/announce-once-ledger.js +24 -2
- package/dist/core/runner/assemble-result.d.ts +6 -0
- package/dist/core/runner/assemble-result.js +1 -1
- package/dist/core/runner/auto-strip-notice.d.ts +36 -0
- package/dist/core/runner/auto-strip-notice.js +59 -0
- package/dist/core/runner/branch-window.d.ts +3 -0
- package/dist/core/runner/branch-window.js +37 -0
- package/dist/core/runner/caller-roster.d.ts +8 -3
- package/dist/core/runner/capability-withheld-seat.js +3 -9
- package/dist/core/runner/clock-and-limits.d.ts +22 -1
- package/dist/core/runner/clock-and-limits.js +30 -0
- package/dist/core/runner/contracts.d.ts +58 -35
- package/dist/core/runner/credential-refusal-seat.d.ts +2 -0
- package/dist/core/runner/credential-refusal-seat.js +47 -0
- package/dist/core/runner/cross-boundary-lane.js +6 -4
- package/dist/core/runner/decide-continuation.js +1 -1
- package/dist/core/runner/file-history-restore.d.ts +53 -2
- package/dist/core/runner/file-history-restore.js +29 -10
- package/dist/core/runner/forward-sink.d.ts +14 -1
- package/dist/core/runner/forward-sink.js +14 -12
- package/dist/core/runner/gate-exit.d.ts +8 -7
- package/dist/core/runner/gate-exit.js +13 -13
- package/dist/core/runner/git-status-frame.js +2 -4
- package/dist/core/runner/initial-run-state.js +1 -1
- package/dist/core/runner/main-call-account.d.ts +3 -1
- package/dist/core/runner/main-call-account.js +5 -3
- package/dist/core/runner/mcp-redial.d.ts +2 -5
- package/dist/core/runner/mcp-redial.js +2 -2
- package/dist/core/runner/nested-memory-content.d.ts +7 -1
- package/dist/core/runner/nested-memory-content.js +73 -13
- package/dist/core/runner/nested-memory-lane.d.ts +44 -19
- package/dist/core/runner/nested-memory-lane.js +152 -76
- package/dist/core/runner/park-commit.js +3 -1
- package/dist/core/runner/permission-rule-lanes.d.ts +8 -2
- package/dist/core/runner/permission-rule-lanes.js +29 -3
- package/dist/core/runner/person-path-rows.d.ts +3 -11
- package/dist/core/runner/person-path-rows.js +21 -24
- package/dist/core/runner/prepare-acquire-reconcile.js +3 -1
- package/dist/core/runner/prepare-artifact.d.ts +1 -1
- package/dist/core/runner/prepare-artifact.js +20 -14
- package/dist/core/runner/prepare-ask-lane.d.ts +9 -6
- package/dist/core/runner/prepare-ask-lane.js +6 -5
- package/dist/core/runner/prepare-boundary-parks.d.ts +7 -1
- package/dist/core/runner/prepare-boundary-parks.js +27 -24
- package/dist/core/runner/prepare-caps-and-workflow.d.ts +5 -2
- package/dist/core/runner/prepare-caps-and-workflow.js +4 -3
- package/dist/core/runner/prepare-config-doors.d.ts +9 -24
- package/dist/core/runner/prepare-config-doors.js +58 -41
- package/dist/core/runner/prepare-context-lane.d.ts +2 -2
- package/dist/core/runner/prepare-context-lane.js +3 -5
- package/dist/core/runner/prepare-defer-classify.d.ts +1 -1
- package/dist/core/runner/prepare-defer-classify.js +1 -1
- package/dist/core/runner/prepare-delegation-surface.d.ts +2 -2
- package/dist/core/runner/prepare-delegation-surface.js +3 -2
- package/dist/core/runner/prepare-file-history.d.ts +62 -26
- package/dist/core/runner/prepare-file-history.js +117 -46
- package/dist/core/runner/prepare-gate-stations.d.ts +4 -0
- package/dist/core/runner/prepare-gate-stations.js +6 -5
- package/dist/core/runner/prepare-hands-readface.d.ts +9 -4
- package/dist/core/runner/prepare-hands-readface.js +8 -4
- package/dist/core/runner/prepare-inherited-gate.d.ts +11 -4
- package/dist/core/runner/prepare-inherited-gate.js +6 -6
- package/dist/core/runner/prepare-listings.d.ts +1 -1
- package/dist/core/runner/prepare-listings.js +1 -1
- package/dist/core/runner/prepare-memory.js +2 -2
- package/dist/core/runner/prepare-offload-wrappers.d.ts +1 -1
- package/dist/core/runner/prepare-offload-wrappers.js +1 -1
- package/dist/core/runner/prepare-park-ask.d.ts +10 -3
- package/dist/core/runner/prepare-park-ask.js +7 -9
- package/dist/core/runner/prepare-policy-chain.d.ts +18 -9
- package/dist/core/runner/prepare-policy-chain.js +18 -13
- package/dist/core/runner/prepare-project-context.d.ts +1 -1
- package/dist/core/runner/prepare-project-context.js +5 -3
- package/dist/core/runner/prepare-prompt-assembly.js +2 -1
- package/dist/core/runner/prepare-prompt-inputs.d.ts +2 -2
- package/dist/core/runner/prepare-prompt-inputs.js +1 -1
- package/dist/core/runner/prepare-protocol-tools.d.ts +4 -5
- package/dist/core/runner/prepare-protocol-tools.js +7 -7
- package/dist/core/runner/prepare-question-face.d.ts +4 -2
- package/dist/core/runner/prepare-run-refs.d.ts +7 -2
- package/dist/core/runner/prepare-run-refs.js +8 -7
- package/dist/core/runner/prepare-safety-scan.d.ts +4 -3
- package/dist/core/runner/prepare-suspend-saga.d.ts +3 -7
- package/dist/core/runner/prepare-suspend-saga.js +4 -3
- package/dist/core/runner/prepare-task.js +66 -53
- package/dist/core/runner/prepare-tool-disclosure-mount.d.ts +3 -2
- package/dist/core/runner/prepare-tool-disclosure-mount.js +2 -1
- package/dist/core/runner/prepare-turn-wiring.d.ts +3 -1
- package/dist/core/runner/prepare-turn-wiring.js +1 -1
- package/dist/core/runner/prepare-wiring-manifest.d.ts +9 -4
- package/dist/core/runner/prepare-wiring-manifest.js +14 -6
- package/dist/core/runner/prepare-workspace-restore.d.ts +4 -4
- package/dist/core/runner/prepare-workspace-restore.js +15 -6
- package/dist/core/runner/remote-env-retry.d.ts +3 -3
- package/dist/core/runner/restore-seat.d.ts +41 -0
- package/dist/core/runner/restore-seat.js +38 -0
- package/dist/core/runner/resume-admission.js +1 -1
- package/dist/core/runner/resume-apply.js +4 -3
- package/dist/core/runner/resume-checkpoint-screen.js +2 -1
- package/dist/core/runner/resume-internals-and-config.js +2 -1
- package/dist/core/runner/resume-preflight.js +2 -1
- package/dist/core/runner/rewind-files.js +22 -6
- package/dist/core/runner/rule-store-unreadable-notice.d.ts +8 -0
- package/dist/core/runner/rule-store-unreadable-notice.js +11 -0
- package/dist/core/runner/run-brain-sinks.d.ts +3 -6
- package/dist/core/runner/run-brain-sinks.js +6 -6
- package/dist/core/runner/run-compaction-machinery.d.ts +2 -2
- package/dist/core/runner/run-git-lane.d.ts +2 -0
- package/dist/core/runner/run-git-lane.js +2 -1
- package/dist/core/runner/run-harness-handlers.js +5 -4
- package/dist/core/runner/run-identity-wiring.js +20 -12
- package/dist/core/runner/run-leg.js +31 -72
- package/dist/core/runner/run-notification-lane.d.ts +3 -0
- package/dist/core/runner/run-notification-lane.js +2 -4
- package/dist/core/runner/run-reasoning-seat.d.ts +5 -6
- package/dist/core/runner/run-reasoning-seat.js +13 -18
- package/dist/core/runner/run-settle-and-teardown.js +13 -9
- package/dist/core/runner/run-terminal-adoption.js +4 -8
- package/dist/core/runner/run-turn-boundary.js +6 -3
- package/dist/core/runner/runtask.d.ts +9 -4
- package/dist/core/runner/runtask.js +29 -22
- package/dist/core/runner/session-rule-policy.d.ts +11 -1
- package/dist/core/runner/shell-env-withheld.js +2 -7
- package/dist/core/runner/side-call-account.d.ts +5 -6
- package/dist/core/runner/side-call-account.js +2 -2
- package/dist/core/runner/stream-lifecycle-verbs.d.ts +3 -0
- package/dist/core/runner/stream-lifecycle-verbs.js +5 -3
- package/dist/core/runner/stream-settle-backstop.js +4 -6
- package/dist/core/runner/stream-steer-verb.d.ts +1 -1
- package/dist/core/runner/stream-steer-verb.js +4 -4
- package/dist/core/runner/sub-agent-deny-lists.d.ts +8 -3
- package/dist/core/runner/sub-agent-deny-lists.js +7 -4
- package/dist/core/runner/teardown-bounded.d.ts +7 -0
- package/dist/core/runner/teardown-bounded.js +7 -0
- package/dist/core/runner/terminal-account.js +41 -13
- package/dist/core/runner/terminal-projection.d.ts +1 -5
- package/dist/core/runner/tool-end-body.d.ts +9 -3
- package/dist/core/runner/tool-end-body.js +6 -3
- package/dist/core/runner/tool-output-projection.d.ts +13 -4
- package/dist/core/runner/tool-output-projection.js +10 -14
- package/dist/core/runner/turn-attachments.d.ts +3 -2
- package/dist/core/runner/turn-attachments.js +12 -5
- package/dist/core/runner-deps.d.ts +6 -6
- package/dist/core/runtime-caps.d.ts +2 -1
- package/dist/core/secret-env.d.ts +20 -3
- package/dist/core/secret-env.js +13 -15
- package/dist/core/session-checkout.d.ts +115 -15
- package/dist/core/session-checkout.js +435 -136
- package/dist/core/session-policy-store.js +21 -3
- package/dist/core/session-reconcile.js +1 -0
- package/dist/core/session-store.js +2 -0
- package/dist/core/session.d.ts +11 -1
- package/dist/core/shell-run-candidates.d.ts +1 -0
- package/dist/core/shell-run-candidates.js +1 -1
- package/dist/core/shutdown-debug.d.ts +6 -5
- package/dist/core/shutdown-debug.js +5 -1
- package/dist/core/skill-grant.js +4 -1
- package/dist/core/skills-directory.js +5 -3
- package/dist/core/spec-contract.d.ts +1 -1
- package/dist/core/store-contracts/checkpoint-store-contract.js +2 -7
- package/dist/core/store-contracts/file-history-store-contract.js +26 -2
- package/dist/core/store-contracts/session-deletion-cascade-contract.js +130 -14
- package/dist/core/store-contracts/task-list-store-contract.d.ts +2 -1
- package/dist/core/store-contracts/task-list-store-contract.js +28 -1
- package/dist/core/task-event.d.ts +48 -27
- package/dist/core/task-notification.d.ts +15 -9
- package/dist/core/task-notification.js +12 -0
- package/dist/core/task-registry-agent.d.ts +24 -22
- package/dist/core/task-registry-agent.js +65 -46
- package/dist/core/task-registry-shared.d.ts +17 -5
- package/dist/core/task-registry.d.ts +17 -6
- package/dist/core/task-registry.js +6 -1
- package/dist/core/task-result.d.ts +71 -25
- package/dist/core/task-spec-members.d.ts +0 -1
- package/dist/core/task-spec-members.js +1 -1
- package/dist/core/task-spec.d.ts +60 -52
- package/dist/core/task-stream.d.ts +45 -12
- package/dist/core/terminal-cause.d.ts +45 -8
- package/dist/core/terminal-cause.js +5 -0
- package/dist/core/tool-card-bound.d.ts +32 -0
- package/dist/core/tool-card-bound.js +33 -0
- package/dist/core/tool-catalog-entries.d.ts +9 -0
- package/dist/core/tool-catalog-entries.js +48 -0
- package/dist/core/tool-errors.d.ts +4 -3
- package/dist/core/tool-face.d.ts +3 -3
- package/dist/core/tool-mount-facts.d.ts +3 -2
- package/dist/core/tool-mount-facts.js +1 -1
- package/dist/core/tool-mount-static.js +7 -5
- package/dist/core/tool-policy.d.ts +21 -19
- package/dist/core/tool-policy.js +5 -6
- package/dist/core/tool-registry.d.ts +3 -1
- package/dist/core/tool-registry.js +15 -1
- package/dist/core/tool-result-store.d.ts +9 -3
- package/dist/core/tool-result-store.js +14 -1
- package/dist/core/tool-roster.d.ts +6 -0
- package/dist/core/tool-roster.js +18 -0
- package/dist/core/tool-spec.d.ts +17 -8
- package/dist/core/trace.d.ts +35 -31
- package/dist/core/types.d.ts +2 -2
- package/dist/core/untrusted-egress.d.ts +10 -9
- package/dist/core/untrusted-egress.js +9 -9
- package/dist/core/untrusted-text.d.ts +0 -18
- package/dist/core/untrusted-text.js +7 -3
- package/dist/core/user-attachment.d.ts +9 -6
- package/dist/core/user-attachment.js +12 -10
- package/dist/core/user-row-origin.d.ts +44 -25
- package/dist/core/user-row-origin.js +30 -20
- package/dist/core/wiring-manifest.d.ts +26 -7
- package/dist/core/wiring-manifest.js +14 -11
- package/dist/core/workspace-snapshot.d.ts +3 -1
- package/dist/core/workspace-snapshot.js +1 -0
- package/dist/engine/execution-env/node-execution-env.d.ts +23 -16
- package/dist/engine/execution-env/node-execution-env.js +127 -67
- package/dist/engine/harness/agent-harness.d.ts +4 -5
- package/dist/engine/harness/agent-harness.js +2 -7
- package/dist/engine/harness/types.d.ts +51 -19
- package/dist/engine/harness/types.js +28 -1
- package/dist/engine/llm/types.d.ts +26 -14
- package/dist/engine/llm/types.js +2 -2
- package/dist/engine/loop/agent-loop.d.ts +2 -2
- package/dist/engine/loop/agent-loop.js +51 -275
- package/dist/engine/loop/types.d.ts +0 -18
- package/dist/engine/session/import-validate.d.ts +3 -0
- package/dist/engine/session/import-validate.js +34 -2
- package/dist/engine/session/repo-utils.d.ts +42 -4
- package/dist/engine/session/repo-utils.js +33 -5
- package/dist/engine/session/session.d.ts +41 -14
- package/dist/engine/session/session.js +96 -25
- package/dist/index.d.ts +21 -16
- package/dist/index.js +16 -10
- package/dist/orchestration/run-spec.d.ts +1 -1
- package/dist/orchestration/run-workflow-tool.d.ts +12 -9
- package/dist/orchestration/run-workflow-tool.js +15 -19
- package/dist/orchestration/workflow-governance.js +13 -21
- package/dist/orchestration/workflow-observe.d.ts +4 -3
- package/dist/orchestration/workflow-primitives.d.ts +1 -1
- package/dist/orchestration/workflow-record-artifact.d.ts +14 -0
- package/dist/orchestration/workflow-record-artifact.js +32 -0
- package/dist/orchestration/workflow-script-store.d.ts +9 -0
- package/dist/orchestration/workflow-script-store.js +20 -1
- package/dist/orchestration/workflow-types.d.ts +33 -3
- package/dist/orchestration/workflow.d.ts +24 -14
- package/dist/orchestration/workflow.js +52 -23
- package/dist/prompt-assembly/artifact.js +4 -8
- package/dist/prompt-assembly/catalog/engine.d.ts +263 -0
- package/dist/prompt-assembly/catalog/engine.js +30 -0
- package/dist/prompt-assembly/catalog/index.d.ts +1 -1
- package/dist/prompt-assembly/catalog/index.js +2 -2
- package/dist/prompt-assembly/catalog/locales/zh-Hans/engine.d.ts +1 -1
- package/dist/prompt-assembly/catalog/locales/zh-Hans/engine.js +29 -0
- package/dist/prompt-assembly/epoch.js +6 -8
- package/dist/prompt-assembly/turn-snapshot.js +8 -9
- package/dist/stores/file/background-agent-store.js +5 -6
- package/dist/stores/file/checkpoint-store.d.ts +0 -4
- package/dist/stores/file/checkpoint-store.js +7 -3
- package/dist/stores/file/file-history-store.d.ts +8 -0
- package/dist/stores/file/file-history-store.js +72 -0
- package/dist/stores/file/fs-atomic.js +17 -12
- package/dist/stores/file/index.js +11 -5
- package/dist/stores/file/mailbox-store.js +13 -4
- package/dist/stores/file/permission-rule-store.js +14 -13
- package/dist/stores/file/shared-ledger.d.ts +4 -0
- package/dist/stores/file/shared-ledger.js +4 -0
- package/dist/stores/file/task-list-store.js +12 -3
- package/dist/stores/file/workflow-run-store.js +1 -0
- package/dist/tools/fs/bash-program-position.d.ts +4 -2
- package/dist/tools/fs/bash-program-position.js +1 -1
- package/dist/tools/fs/bash-readonly-classifier.js +1 -1
- package/dist/tools/fs/bound-open.d.ts +97 -0
- package/dist/tools/fs/bound-open.js +103 -0
- package/dist/tools/fs/containment-face.d.ts +2 -2
- package/dist/tools/fs/fs-pdf.d.ts +7 -1
- package/dist/tools/fs/fs-pdf.js +31 -23
- package/dist/tools/fs/fs-read.d.ts +24 -17
- package/dist/tools/fs/fs-read.js +84 -53
- package/dist/tools/fs/fs-shared.d.ts +14 -12
- package/dist/tools/fs/fs-shared.js +5 -4
- package/dist/tools/fs/fs-write.d.ts +2 -2
- package/dist/tools/fs/fs-write.js +20 -12
- package/dist/tools/fs/index.d.ts +4 -4
- package/dist/tools/fs/index.js +2 -2
- package/dist/tools/fs/notebook.d.ts +1 -1
- package/dist/tools/fs/pdf.d.ts +12 -3
- package/dist/tools/fs/pdf.js +22 -9
- package/dist/tools/fs/read-deny-home.d.ts +3 -2
- package/dist/tools/fs/read-deny-home.js +8 -11
- package/dist/tools/fs/safety.d.ts +3 -2
- package/dist/tools/fs/shell-read-check.js +10 -9
- package/dist/tools/monitor.d.ts +6 -5
- package/dist/tools/task-list.d.ts +16 -6
- package/dist/tools/task-list.js +4 -0
- package/dist/tools/web.d.ts +9 -1
- package/dist/tools/web.js +8 -3
- package/dist/tools/worktree.d.ts +7 -5
- package/dist/tools/worktree.js +4 -3
- package/package.json +4 -3
- package/test/export-surface.snapshot.json +93 -7
- package/wire-changes/core-stores-sql@7.68.0.json +39 -0
- package/wire-changes/core@7.68.0.json +4257 -0
- package/wire-changes/mods-kernel@7.68.0.json +129 -0
- package/wire-changes/primitives@7.68.0.json +253 -0
- package/wire-changes/wiki-gate@7.68.0.json +39 -0
- package/dist/core/runner/credential-origin-seat.d.ts +0 -2
- package/dist/core/runner/credential-origin-seat.js +0 -20
package/CHANGELOG.md
CHANGED
|
@@ -1,3 +1,1519 @@
|
|
|
1
|
+
## 7.68.0 — 2026-11-13
|
|
2
|
+
|
|
3
|
+
### `incarnationCell(existing, tokenless?)` can give a row that holds no token an unmarked one — a host giving a token to an old row no longer has to pick between no token and one that moves the row's file-history scope to an address (#4394; additive: one optional parameter, the default is byte for byte today's; behavior change: none for any call that passes one argument; narrowing: none; widening: none; hard-breaking: none; closed sets: none on the wire (the parameter takes two words); exports: none added, none removed; wire: none)
|
|
4
|
+
|
|
5
|
+
- **Before.** The mint point answered a row with no token `{}` and a new row a token with the address-epoch mark. A host whose own occupant test needs every row to hold a token, giving one to a row that already holds engine state, could not get an unmarked one from the package: `incarnationCell(undefined)` marks it, and a marked token moves the row's file-history scope from its bare id to an address — the history recorded at the bare id is then not read by a rewind and not reached by the host's erase.
|
|
6
|
+
- **Now.** `incarnationCell(existing, tokenless = "keep")`. No row ⇒ a marked token; a row holding a token ⇒ that token; a row with none × `"keep"` ⇒ `{}` (today's answer); a row with none × `"mint"` ⇒ a random v4 without the mark, so `fileHistoryScopeOf(id, token) === id`. `tokenless` plays no part in the first two. Any word outside the two is refused on every row, `config.invalid_argument`, naming the two words and never the value.
|
|
7
|
+
- **Not changed.** Every bundled store's call (all pass `undefined` and no second argument), the bytes of every answer for a call with one argument, `fileHistoryScopeOf`, `readFileHistoryScope`, and where any row's history lives.
|
|
8
|
+
- **Known limit (KL-4394-a, #4404).** Do not give `"mint"`'s token to a row whose session already holds a session tree of its own: the tree's birth mark was written naming no occupant, and once the row holds a token the mark reads as another occupant's (`src/core/session-checkout.ts`, `recordHoldsMark`), so the session's next leg is refused `workspace.checkout_missing` — measured by the release review leg, the same write-back with `"keep"` runs. The integration guide's paragraph on giving an old row a token now says so.
|
|
9
|
+
- **Rules.** One mint point, one more parameter; no second place mints. Retired: none. Added: none.
|
|
10
|
+
- **Pins.** `test/backlog4394-incarnation-cell-tokenless-mint.test.ts` — the four cells, the refusal on every kind of row, a thousand mints none marked, and the real path on the bundled file repo and file history store: an old row with history at its bare id, a `"mint"` token written back keeps the scope and the history, the same write-back with `incarnationCell(undefined)`'s token (the witness) moves the scope and the history is no longer read.
|
|
11
|
+
|
|
12
|
+
### The frame a resumed leg opens with for an approved tool call says the approval and points at the tool results — it no longer says the call "has now been executed" (#4007; server S-3598 R189; behavior change: one model-facing sentence reworded; narrowing: one, below, open to veto @server @cli @test; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none (the sentence rides a user message, not a typed member)) (@server @cli @test)
|
|
13
|
+
|
|
14
|
+
- **Before.** The continuation frame of a resumed approval read `The pending tool call "X" was APPROVED and has now been executed — its result is in the tool results above.` It was built after the approved call was resolved, from the decision alone, so it said "executed" whatever had come of the call: a call that ran and failed (its execution threw), and a call stopped before it was let go (an edit its tool could not read, an approval object that no longer matched), both read as executed on the line the model reads right after the result row that says otherwise.
|
|
15
|
+
- **Now.** `The pending tool call "X" was APPROVED — what came of the call (its result, or why it did not run) is in the tool results above.` One sentence for every outcome, a function of the decision alone: the approval is the fact the frame holds, the call's own result row says whether it ran and how it ended. The rest of the frame (the resumed-task order, the carry-on clause, a refusal's sentences) is byte for byte what it was.
|
|
16
|
+
- **Narrowing (open to veto @server @cli @test).** The frame no longer asserts that the approved call executed. A reader that matched `has now been executed` on this frame reads the new words; none in this repository's source or docs does (the frame's old words were pinned in one test file and read nowhere else), a consumer's own match is not read from here.
|
|
17
|
+
- **Not changed.** What runs on an approval and in which order (the call is resolved before the frame is built); the result row; the refusal frames (a person's, a policy's, an elapsed window's); the plan-review, dry-run and wake frames.
|
|
18
|
+
- **Rules.** One sentence for one decision where the outcome used to be implied; none added, none retired as a rule (one spelling changed). Added: none.
|
|
19
|
+
- **Pins.** `test/backlog4007-approved-continuation-says-approval-only.test.ts` (G1 an approved call whose execution throws, G1b an approved call never let go, G2 the control: an approved call that runs reads the same bytes as the other two and keeps the approval, the pointer and the resumed-task order, G3 the sentence as a pure function of the decision in both postures and the machine refusals unchanged); `test/backlog1624-plan-rejected-stop-and-wait.test.ts` H5 no longer pins the approve arm's old words.
|
|
20
|
+
- **Consumers.** @server (S-3598 residual, R189): the frame no longer claims the execution a parked sub-agent's approval did not get; the outcome is the result row's. @cli: a card or transcript line built from this sentence's old words reads the new ones. @test: a black-box read of the resumed leg's first model call (a parked call approved, then failing or refused before it ran) finds no execution claim in the frame.
|
|
21
|
+
|
|
22
|
+
### A question card that was shown and then ran out its window is told to the model as an unanswered question, never as "NOT shown" — the question seam's contract now says which word a face that rendered the card owes (#3986 (b); documentation only — no new outcome arm, no new `OnQuestionOutcome` form; open to veto @server)
|
|
23
|
+
|
|
24
|
+
- **Rule.** A face that puts a question in front of a person owes the engine exactly one of three things: an answer; `{ kind: "unavailable" }` — nothing was shown, zero interaction; or a throw — the card was, or may have been, shown and no answer came, an elapsed window included. The engine maps a throw to the delivery-failure arm, which never re-routes the question; only `unavailable` re-routes it, and the refusal of that arm says that no live face answered it (#4012).
|
|
25
|
+
- **Before.** The `QuestionUnavailable` contract (`src/core/ask-question.ts`) said a face that showed the card and then failed must throw, but not what returning `unavailable` after rendering is. A host whose window elapsed on a card a person had seen and that answered `unavailable` (server S-3395, a synchronously delegated child's question) had the model told "NOT shown" about it: a false sentence whose cause is the host's return, not an engine classification gap.
|
|
26
|
+
- **Now.** The contract names it: returning `unavailable` once the card was shown, a window that elapsed with nobody answering included, is a violation of the contract. No engine change. A face that throws on an elapsed window reaches the model as the headless continuation — "No human is available to answer right now. Proceed with your best judgment: pick the most reasonable option, state the assumption you made, and continue. (no answer was obtained: <the face's own thrown text>)" — or, under interaction posture `interactive`, the coded failure `question.human_channel_failed` — "Error (AskUserQuestion): the wired human channel failed to deliver an answer (<the face's own thrown text>). This run declared interaction posture "interactive", so the engine will not silently self-answer; fix the question channel (or opt into the synthetic continuation explicitly via interactiveQuestionFallback)." Neither says "NOT shown", and the host's own account of the elapsed window rides inside. The call's gate record on that path is `{ disposition: { kind: "allowed" } }` with no settlement and no origin (the delivery-failure arm has no settlement word, and none is added).
|
|
27
|
+
- **Not changed.** `ContentAskOutcome` (`answered` / `unavailable` / `delivery_failure`), `OnQuestionOutcome`, `SettlementKind` and `SETTLEMENT_NOBODY_JUDGED`, and every sentence the model reads on the delivery-failure arm; an `unavailable` outcome with no park armed still ends in the refusal (the zero-interaction case, kept as the control; its words are #4012's).
|
|
28
|
+
- **Rules.** Retired: none. Added: none — one sentence in the existing contract. The false-sentence reading it removes goes 1 → 0 for a host that follows the contract.
|
|
29
|
+
- **Consumers.** @server (S-3395): the question face of a synchronously delegated child, on a rendered card whose window elapsed, THROWS (the elapsed-window text is the thrown message; the model reads it) and never answers `unavailable`; core adds no third arm. @cli @client-core: none. @test: G1.
|
|
30
|
+
- **Wire.** None: no key, no word, no frame moves.
|
|
31
|
+
- **Pins.** `test/backlog3986-shown-then-expired.test.ts` — G1-a headless: the face is handed the card, throws an elapsed-window message; the run completes, no checkpoint is written, the model's sentence carries the face's text and not "NOT shown", the gate record is an allow with no settlement · G1-b the same under posture `interactive`: the in-band coded card `question.human_channel_failed`, no "NOT shown" · G2 control: a face that answers `unavailable` with no park armed ⇒ the refusal ("no live face answered it", #4012) · G3 the contract sentence names the violation. G3 is red on the base tree; G1 and G2 are green there by design (regression pins for the reading the contract already had). Mutations (each red where it should be): a thrown face folded into `unavailable` ⇒ G1-a / G1-b red; the refusal's clause dropped ⇒ G2 red; the contract sentence removed ⇒ G3 red; the headless sentence saying "NOT shown" ⇒ G1-a red, the interactive one ⇒ G1-b red (each alone).
|
|
32
|
+
|
|
33
|
+
### On the question seam "shown" means presented on a live question face — a ledger or inbox delivery with no live face is not a presentation, and the refusal of a question nothing can carry says "no live face answered it" instead of "NOT shown to anyone" (#4012; server S-3764; behavior change: one model-read refusal sentence reworded, one contract sentence added; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none; wording change open to veto @server @cli @client-core @test)
|
|
34
|
+
|
|
35
|
+
- **Rule.** One definition, in the seam's contract: "shown" is presented on a live question face — a face a person can answer on now. A face that delivers the question to a ledger or an inbox with no live face has shown nothing, so `unavailable` is the right return for it and the engine re-routes the question (park it for a later human, or refuse it). A face that presented the card on a live surface and whose window then elapsed throws, as before; returning `unavailable` for that stays the violation. The refusal a question meets when nothing can carry it says the one thing true of every way of reaching it.
|
|
36
|
+
- **Before.** The contract said a card was "shown" without saying where, and the refusal told the model "the question was NOT shown to anyone": a claim about the whole delivery where only the live presentation was meant (a deployment whose face wrote the question to its ledger and answered `unavailable` had the model told the question was not shown to anyone, while a person could still read it there), and false for a card a live face had rendered before it reported nobody home.
|
|
37
|
+
- **Now.** `QuestionUnavailable`'s contract carries the definition above and the ledger / inbox clause; the reading of the earlier sentence (a card "shown" then answered `unavailable` is a violation) is the live-face one. The refusal now reads `AskUserQuestion could not be parked for a durable human answer on this path (no checkpoint facility, or the suspend failed) and a synchronous permission approver cannot carry an answer — no live face answered it. Proceed with your best judgment and state the assumption you made.` — one sentence for both settlement words the refusal files (`no_approver`: no content channel at all; `approver_unavailable`: the channel answered `unavailable`, or was kept off the face by a durable mandate).
|
|
38
|
+
- **Wording (open to veto @server @cli @client-core @test).** A reader that matched `NOT shown to anyone` on this refusal reads the new closing clause; none in this repository's source or docs does beyond the pins re-pinned below, a consumer's own match is not read from here.
|
|
39
|
+
- **Not changed.** The two settlement words, the three-valued outcome (`OnQuestionOutcome`: an answer, `unavailable`, a throw), the park re-route of an `unavailable` outcome, the delivery-failure arm and every sentence the model reads on it, and what the engine does with a thrown face.
|
|
40
|
+
- **Rules.** One false claim retired (the refusal's "NOT shown"); one definition added once, in the contract, which the engine's comments and the refusal point back to; no third state.
|
|
41
|
+
- **Pins.** `test/backlog4012-shown-is-a-live-face.test.ts` — R1 no park armed, a face that reports nobody reachable: "no live face answered it", never "NOT shown", settlement `approver_unavailable` · R2 a durable mandate over a store that fails to park, the face never consulted: the same words · L1 control: a face that writes the question to its ledger and answers `unavailable` ⇒ one park, a paused run, nothing thrown · C1 the contract defines "shown" as presented on a live question face, names the ledger / inbox delivery as not shown, keeps the violation. R1, R2 and C1 are red on the base tree, L1 green by design (the engine's reading was already that). Re-pinned to the new words: `test/backlog3986-shown-then-expired.test.ts` G2 (and G1's negations name the refusal's clause too), `test/question-sync-first.test.ts`, `test/auto-mode.test.ts`, `test/backlog522-durable-gate-loud.test.ts`. Mutations, each read per cell: the old sentence back ⇒ R1, R2 and the re-pinned positives red; `live` dropped from the sentence ⇒ R1, R2 red; the park re-route of `unavailable` removed ⇒ L1 red; the ledger clause dropped from the contract ⇒ C1 red; the definition dropped ⇒ C1 red.
|
|
42
|
+
- **Consumers.** @server (S-3764, judged by the live face): a face with no live surface writes to its ledger and answers `unavailable` (the engine parks the question); a face that rendered the card on a live surface throws on an elapsed window. @cli @client-core: nothing new to render; a match on the old closing clause reads the new one. @test: the black-box cells that read "NOT shown" on this refusal read "no live face answered it".
|
|
43
|
+
|
|
44
|
+
### The bind of a background-agent row's child session refuses a session id that carries an invisible character anywhere in it, and a task id that names no background-agent row — one judge, every refusal `config.invalid_argument` (#2997; bad-value rule; narrowing — open to veto @server)
|
|
45
|
+
|
|
46
|
+
- **Rule.** A bad value is refused where it is read, loudly. The bind of a child session onto its row has ONE judge and ONE mint for what it refuses: the session id (a string with no separator, control or format character anywhere in it), the task id (a background-agent row that exists), and the row's state (not anchored to another session already). A value is refused as it came: nothing is trimmed or rewritten (a rewritten id would be a second source of the anchor).
|
|
47
|
+
- **Before.** `TaskRegistry.bindBackgroundAgentSession(id, sessionId, store)` refused only a non-string or empty session id (#2963) and a re-anchor. A blank id (`" "`), one padded with white space (`" child-ok "`), one carrying a control character (a CR reads as another line wherever the anchor is printed), one with a format character between visible letters (a right-to-left override) and a task id that names no background-agent row were taken — the last in silence: the bind returned, so a caller that had lost its row believed it had anchored a session. The refusals it did have were plain `Error`s with no code.
|
|
48
|
+
- **Now.** `bindBackgroundAgentSessionLane` (`src/core/task-registry-agent.ts`) judges in one place and mints every refusal through `bindRefusal` as `config.invalid_argument` (`.code` and `details.code`, the shape a tool body's coded refusal has): a session id that is not a string, is empty, or carries an invisible character anywhere in it — a separator (`\p{Z}`: a space, a no-break space, a line or paragraph separator), a control (`\p{Cc}`: C0, DEL, C1) or a format character (`\p{Cf}`: zero width, a direction override, a byte order mark) — is "a background-agent session anchor must be a non-empty session id with no invisible character — no separator, control or format character (got <what it is>)", never the value; the id is refused whole, so a padded `" child-ok "` is refused and not stored as `child-ok`; a task id that names no background-agent row — gone, never registered, or a row of another kind — is "no background-agent row "<id>" to anchor a session to", the id JSON-quoted (a CR prints as its escape) and cut at 80 characters; a row anchored to another session keeps refusing in the words it had, now coded. Judged in that order, the session id first (a bad id on a missing row names the id).
|
|
49
|
+
- **Narrowing (open to veto @server).** A bind for a row that is gone, or that is not a background-agent row, used to return silently and now throws. The three engine callers (`src/agents/subagent.ts`: the fork lane and the agent lane's two binds) bind a row they registered in the same turn, and the registry's sweeps evict only terminal rows past their ttl, so a running row is never gone under them; a host that calls the verb with a stale id gets the refusal it was owed (the one case that can still reach a vanished row is KL-2997-a). An id with a space or a format character anywhere in it is refused too; the engine's own ids (uuid-shaped, the store's mint) carry none. There is no length bound: an id of any length made of visible characters anchors (a bound with no reason would be a special case). The #2963 pin that said "a row that is gone stays silent" is re-pinned.
|
|
50
|
+
- **Not changed.** The same id again is still a no-op with no second durable write; release then bind; the empty-string and non-string words (`(got an empty string)` / `a number` / `a null`); a valid id of any other shape (uuid, punctuation, non-ASCII letters, a combining mark) still anchors.
|
|
51
|
+
- **Rules.** Judges of the bind's arguments 2 → 1 (the inline #2963 check and the would-be second one for the new values are one function); readings of "an id that is no anchor" 3 (blank, control, and the two forms a second pass found — padded, a format character inside) → 1 (one character class, anywhere in the string); refusal mints 3 (one uncoded `Error` per refusal) → 1.
|
|
52
|
+
- **Consumers.** @server @cli @client-core: none (an internal registry verb; no wire frame carries the refusal).
|
|
53
|
+
- **Pins.** `test/backlog2997-bind-bad-values.test.ts` — X1 five spellings of nothing-but-invisible (spaces, tab, no-break space, zero-width space, lone joiner) refused as a string with an invisible character, the row unanchored on every face and no durable write spent · X2 five control spellings (CR, LF, NUL, DEL, C1) refused in the same word, the value never echoed; an ordinary-id control (uuid-shaped, punctuation, non-ASCII letters, a combining mark) still anchors · X5 five separator forms around or inside a visible id (padded with spaces, a trailing tab, a space inside, no-break spaces around, a line separator inside): refused whole, the row not anchored to a trimmed copy, the clean spelling binds afterwards · X6 four format forms (a right-to-left override, a zero-width space and a joiner between letters, a byte order mark in front) · X3 an unknown id, an id with a CR (printed by its escape), a row of another kind, and the order of the two judgments · X4 six refusals (blank, control, non-string, empty, unknown row, re-anchor) all `config.invalid_argument` on `.code` and `details.code`, none touching the anchor; `test/backlog2963-bind-session-anchor-law.test.ts` B3 re-pinned. On the base tree 25 of the 37 pins in the two files are red (24 of the 25 new ones, X2's ordinary-id control green, and B3). Mutations, per fixture: the class check removed ⇒ 20 red (all of X1, X2, X5, X6 and X3-d); the class cut to `\p{Cc}` ⇒ 13 red (the tab and trailing-tab fixtures stay green); without `\p{Cf}` ⇒ 6 red (the zero-width space and joiner in X1, all of X6); without `\p{Z}` ⇒ 7 red (spaces and no-break space in X1, the four separator X5 fixtures other than the trailing tab, X3-d); trim before judging ⇒ 8 red (the white-space fixtures of X1 and X5, the byte order mark in X6, X3-d); unknown row silent ⇒ X3-a, X3-b, X3-c, X4 and B3 red; uncoded mint ⇒ X4 red; re-anchor uncoded ⇒ X4 red.
|
|
54
|
+
|
|
55
|
+
### The deployment credential's refusal is announced to the person like the per-model credential's — one notice reading for the two refusals that keep a credential from a host it was not given for: `route.credential_mismatch` (#3002; user-visible consistency; hard-breaking: one notice word; open to veto @cli @server @client-core)
|
|
56
|
+
|
|
57
|
+
- **Rule.** The request gate has two refusals whose job is to keep a credential from a host it was not given for: the per-model credential bound to another origin (`route.credential_origin_mismatch`) and the deployment credential paired with another root (`route.credential_mismatch`). The person whose request was refused hears about either, once per run, in one account.
|
|
58
|
+
- **Before.** Only the origin refusal was announced (audience `user`, once per run per model and the two origins). The deployment credential's refusal — a catalog entry moved to a host the brain's key was never paired with — reached the person as the request's failure text alone (test c2880's L0 reading of a moved entry under a brain key, case J4: the notices were `config.models_swapped` alone, then the failed terminal), while the sibling refusal of the same gate, for the same kind of move, said it twice.
|
|
59
|
+
- **Now.** `routeRefusalError` reports both refusals to the run's credential-refusal sink (`CredentialRefusalReport`, one union over the two codes, `src/brain/status-sink.ts`) and the one seat, `createCredentialRefusalScope` (`src/core/runner/credential-refusal-seat.ts`), announces each: `route.credential_mismatch`, audience `user`, minter `engine`, once per run per (model, request root, declared root), `detail: { modelId, requestBaseUrl, configBaseUrl, sessionId, runId }` (display spellings; `configBaseUrl` is `""` when the brain declares none; never a credential, never a digest of one), the message naming the model and both roots and the way out (give the model its own credential, or serve it through a brain whose root is its own). Said for a request the gate refused only: a pre-flight or read-face judgment announces nothing, and a derived seat refused this way still falls back as `route.fallback_to_primary` names. `docs/INTEGRATION-CORE.md` §1.10b says the one reading in a sentence.
|
|
60
|
+
- **Not changed.** Both refusals' codes, fix hints and failure text; `RouteRefusalCode` and the pairing statuses; `route.credential_origin_mismatch`'s own message, detail and dedup unit; the other route refusals (`route.credential_missing`, `route.credential_answer_conflict`, `config.protocol_plaintext_credentials`) keep no notice — they keep no credential from any host (nothing would have been sent), the boundary of the family.
|
|
61
|
+
- **Hard-breaking.** (1) *What changed:* `EngineNoticeCode` gains `route.credential_mismatch` (`ENGINE_NOTICE_CODES` +1; the release's count, 135 → 153, is in the #3832 section). *Who is hurt:* a consumer that keys a record by `EngineNoticeCode` — a notice allowlist or router, an audience table, a renderer — gets a compile error until it gives the word a row. *Migration:* add the row (user audience; render `message`, or the roots from `detail`); a surface that already shows the request's failure text may fold the notice into it by code. (2) *What changed:* the module `src/core/runner/credential-origin-seat.ts` is `credential-refusal-seat.ts` and its internal sink names lost "Origin" (`runWithCredentialRefusals`, `reportCredentialRefusal`, `CredentialRefusalReport`, which now carries a `code`). *Who is hurt:* a deep import of those internal paths (none is a package-root name; the wire-changes manifest does not list them). *Migration:* none for a root consumer.
|
|
62
|
+
- **Rules.** Disclosure readings across the family 2 → 1 (one seat, one audience, one dedup law); the added word is the family's second member, not a second rule.
|
|
63
|
+
- **Closed sets.** `EngineNoticeCode` +1 (three tables in `src/core/governance-codes.ts`: the code array, the audience table, the minter table).
|
|
64
|
+
- **Consumers.** @cli @server @client-core: give `route.credential_mismatch` a row in the notice table (user audience; the same renderer as `route.credential_origin_mismatch` reads it: `message`, and the roots from `detail`). @test: c2880 J4 now reads exactly one `user` notice beside the failed terminal.
|
|
65
|
+
- **Pins.** `test/backlog3002-credential-refusal-disclosure.test.ts` — D1 real Runner, real OpenAI brain, two real local listeners: the brain's key is paired with A, the entry points at B with no hook ⇒ zero requests anywhere, the run fails on the code, exactly one `user` notice naming the model and both roots, no credential or digest in any rendering, no origin-word notice · D1b a brain that declares no root ⇒ empty `configBaseUrl`, a sentence that still reads · D2 one refusal repeated is one notice, another request root or declared root is another, the origin refusal in the same scope is its own word, both codes through the one seat · D2b a second run says it again · D2c both words share audience `user` · D3 `adjudicateModelRoute` and the gate: the law announces nothing, the gate's refusal once · D4 `route.credential_missing` announces nothing (the boundary); `test/backlog1821-credential-origin-binding.test.ts` C12 re-pinned to the sink's union. Control: 96 of 96 green across the two files. Mutations, each read per fixture: the mismatch report removed (the base-tree behavior) ⇒ D1, D1b, D2, D2b, D3 red; the dedup key reduced to the model ⇒ D2 red; the origin report removed ⇒ D2 and ten #1821 cases (C12 among them) red; the audience row moved to `operator` ⇒ D1, D2c red; the sentence without the request's root ⇒ D1, D1b red; the latch once per process ⇒ D2b and #1821 W2 red.
|
|
66
|
+
|
|
67
|
+
### A path the engine reads for a host is read as spelled — the worktree-isolation guard, a mod package directory, a skills root and a dangling link's relative target no longer fold a `..` after a symlink before asking the kernel (#4050; behavior change: four readers; narrowing: two, below, open to veto @server @cli @test; widening: one, below, open to veto @server; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @test)
|
|
68
|
+
|
|
69
|
+
- **Rule.** A `..` after a symlink is the filesystem's to resolve. A reader of a path somebody handed it hands the spelling to the kernel as given (`physicalPathOf` over the unfolded spelling) or anchors a relative part at the directory that holds it; it does not fold first with `path.resolve` / `join`, which reads another directory (`<elsewhere>/link/..` is `<elsewhere>`, the kernel's answer is the parent of where `link` points).
|
|
70
|
+
- **Before.** (1) The worktree-isolation guard compared `physicalPathOf(path.resolve(cwd))` with the shared root and the parent's cwd: a factory env whose cwd was spelled `<elsewhere>/link/..` over the shared root (the environment keeps its cwd unfolded, so every I/O landed in the shared root) was read as `<elsewhere>` and accepted. (2) `loadModPackage` resolved its directory with `path.resolve` first. (3) `createSkillsFromDirectory` listed the kernel's directory and named every entry under the folded twin. (4) The memory layout's landing of an act through a dangling link anchored a relative target at the spelled directory, so a link reached through another link landed at the lexical neighbour.
|
|
71
|
+
- **Now.** (1) A worktree child whose factory env is, in the kernel's reading, the shared root or the parent's cwd is refused (`isolation.unavailable`, "shared base root" / "parent's cwd") however the cwd is spelled. (2) The package at `<dir>` is the package the kernel's `<dir>` names. (3) The skills root is resolved once and the entries are listed and read under that one directory. (4) A followed link's relative target is anchored at the directory that holds the link and handed on unfolded, and the missing tail is rejoined under the physical form of the deepest directory that exists — the landing is the kernel's (`landsOutsideStore` and the memory roots' identity follow). A target that steps `..` out of a directory that is not there no longer gets a guessed landing: it throws `ENOENT` (the kernel cannot place that act either); a target ending in a separator is followed as before.
|
|
72
|
+
- **Narrowing (open to veto @server @cli @test).** A host or test whose factory env cwd is spelled `X/link/..` over the shared root or the parent's cwd now fails the isolation guard where it passed; a memory act through a dangling link whose target steps `..` out of a missing directory is refused as a fault (`ENOENT`) where it was given a guessed landing. A path that cannot be canonicalized is refused as before (fail-closed); that now includes the empty string, which `path.resolve` read as the process directory (an env cwd or `rootPath` of `""` fails the guard, `loadModPackage("")` throws `config.mods_invalid`).
|
|
73
|
+
- **Widening (open to veto @server).** A package directory or skills root a host spells with a `..` after a link now loads the kernel's directory, not its lexical twin; a host that relied on the twin sees the other directory's content, or a refusal when that directory holds no package.
|
|
74
|
+
- **Not changed.** The memory surface's and the workflow script store's normalized-spelling reading (KL-902-b), the memory re-anchor destination, a mod manifest's module names, an LSP document path's `..`, the artifact stub's version files (its published paths carry no dot segment) — each stays, and KL-4050-a says why. Wire, exports, closed sets.
|
|
75
|
+
- **Rules.** Retired: the lexical pre-fold at four readers (`prepare-task.ts` canonicalization, `realRoot`, the skills root listing/naming split, the dangling-link anchor). Added: none. Readings of "this spelled directory": 2 → 1 at each.
|
|
76
|
+
- **Pins.** `test/backlog4050-sibling-path-reading.test.ts` — S1 a cwd `<elsewhere>/link/..` over the shared root is refused, with controls (plain spelling refused, own directory accepted, an own `link/..` accepted) · S2a a package directory spelled `<twin>/link/..` loads the kernel's package · S2b a manifest module name stays lexical (boundary) · S4 the published-path grammar refuses dot segments · S5 a skills root spelled `<twin>/link/..` lists and reads one directory · S9 an act through a dangling link inside a linked directory lands at the link target's directory, with controls. The `link/..` cells need a kernel that resolves `..` after a link the POSIX way; the Linux reading is left to the full-run host's run (macOS was measured here).
|
|
77
|
+
- **Consumers.** @server @cli: a deployment factory must root a worktree child's env in a directory that is not, by the kernel's reading, the shared root or the parent's cwd; spell host directories (mod packages, skills roots) without a `..` after a link or expect the kernel's directory. @test: the S1 cell is the black-box shape (factory env cwd `X/link/..`).
|
|
78
|
+
|
|
79
|
+
### A rules value's `grammar` stamp is read as the value's own data member — an accessor, an inherited or a non-enumerable stamp is refused `invalid_shape` by the one reader every station takes — and the one rule for the grammar a value is read in is stated where a snapshot is declared (#3762, #3761; behavior change: narrowing — one, below and open to veto; the reading of a snapshot is stated as it stands, no behavior; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @client-core)
|
|
80
|
+
|
|
81
|
+
- **Rule.** (#3762) The stamp of a rules value is its own enumerable data member, looked at off the property descriptor (`stampMemberOf`): an accessor is never called, and a member a prototype supplies or a non-enumerable one is no stamp of the value's own. One reader serves the write judge's declaration check (`refuseUndeclaredGrammar`), the shape reading (`grammarOf`: `readSessionRulesShape`, so `normalizeRules`, `loosenReasons`, a store's read-back and the run-time layer's snapshot) and the operator part's own stamp. A stamp that is not an own data member is `invalid_shape` — `detail: { field: "grammar", got: "an accessor" | "an inherited member" | "a non-enumerable member" }`, the sentence says "grammar must be an own data member" and quotes nothing a getter answered; an own enumerable member holding `undefined` is still no declaration (`grammar_required`, as before). (#3761) The grammar a value is read in is one rule, now a contract sentence (`session_rules.snapshot_grammar`, on `SessionRuleSource.seed`, and the integration guide §5.3): the value's own stamp; with none, a value beside a store `rev` of 1 or more is a row stored before the stamp (grammar 0), any other value (no `rev`, or the `rev: 0` record of an unwritten key) is read in the grammar of the engine that enforces it. Only a writer declares, because a row outlives the engine that wrote it.
|
|
82
|
+
- **Before.** (#3762) The stamp was read as `value.grammar` in three places, so a getter ran inside the write judge (and again in the shape reading, answering as it liked each time), a member a prototype supplied counted as the writer's declaration — a polluted `Object.prototype.grammar` declared a grammar for every value that carried none — and a non-enumerable one was "declared" although the buckets are read off the value's own enumerable members. (#3761) The reading of a no-stamp, no-`rev` literal handed to `createSessionRulePolicy` was written in a parenthesis of the guide and a "Not changed" line of the #1810 section, and nowhere a reader of the code would meet it.
|
|
83
|
+
- **Narrowing (open to veto @server @client-core).** A rules value whose `grammar` is an accessor, an inherited member or a non-enumerable member used to be accepted (the getter's answer, the prototype's value, taken as the stamp) and is refused `invalid_shape` on every reader of rules: the write judge, `readSessionRulesShape` (a row a store hands back whose stamp is such a member reads as a fault; a snapshot handed to the run-time layer refuses every call, the getter never run), `normalizeRules` and `loosenReasons`. A row built from JSON or a column holds its stamp as an own data member and is unaffected; a host whose store hands rows as class instances with `grammar` on the prototype must hand an own data member (storing the judge's answer, the documented form, already does).
|
|
84
|
+
- **Not changed.** (#3761) Every reading of a value with no stamp: a snapshot with no `rev` is read in the engine's grammar, one beside a `rev` of 1 or more in grammar 0, the `rev: 0` no-row record in the engine's grammar; the write judge still refuses a body that declares none (`grammar_required`). Option (a) of the ticket — demanding a declaration of a snapshot at `createSessionRulePolicy` — was not taken: it would add a special case at one door while `readSessionRulesShape`, `normalizeRules` and `loosenReasons` (root exports the server's `session-sync.ts` reads through) keep the rule, and no host calls the layer with a literal in production (read from origin/main: server `src` 0, client-core `src` 0, cli 0; tests and probes only).
|
|
85
|
+
- **Known limits.** KL-3762-a: the stamp is the one member this reading takes off a descriptor; the rest of a body is still read off the live value — an own accessor bucket (or `rev` / `gen`) runs its getter in `Object.entries`, and a Proxy body runs its traps — a body receipt of the whole value is a larger change than this one.
|
|
86
|
+
- **Rules.** −1 +1: the stamp had three live reads (`grammarOf`, the declaration check, the operator part's unstamped test) and now has one (`stampMemberOf`); the grammar a value is read in was written as a parenthesis and a changelog line and is now one contract sentence with one home. No new refusal word: `invalid_shape` and its `detail` shape are the ones the stamp's own faults already answered with.
|
|
87
|
+
- **Pins.** `test/backlog3762-grammar-stamp-own-data.test.ts` (base column `edda11c7c`: 11 red / 5 green controls) — P0 plain, frozen, no-prototype and own-stamped class-instance bodies accepted · P1 five forms (own getter, throwing getter, prototype member, prototype getter on a class instance, non-enumerable) refused on the judge and on both stores, operator and not, no getter call, nothing written; a polluted `Object.prototype.grammar` declares nothing · P2 `conflict` first, an own `undefined` stamp stays `grammar_required`, a getter answering `undefined` is a getter, the bad-stamp sentences unchanged · P3 the shape reading (a row beside a `rev`, the operator part with its path) and the run-time layer read it the same way · P4 the sentence names the member and quotes no answer. `test/backlog3761-snapshot-grammar-reading.test.ts` (holds on the base — no behavior moves) — S1 the stamp wins beside any `rev` · S2 beside a `rev` of 1 or more: grammar 0, announced · S3 no `rev` and `rev: 0`: the engine's grammar, a constant source's seed alike · S4 the write judge refuses the same literal · S5 `loosenReasons` reads the same grammars.
|
|
88
|
+
- **Consumers.** @server: nothing to change unless a store hands rows whose stamp is not an own data member (a store that keeps the judge's answer and serves it back through `readSessionRulesShape(body).rules` does not). @client-core: its two consumption probes hand `createSessionRulePolicy` an unstamped literal and keep their reading (the engine's grammar); a probe that wants it pinned across an upgrade adds `grammar: SESSION_RULES_GRAMMAR`.
|
|
89
|
+
|
|
90
|
+
### The WebFetch summarizer asks for thinking off on a wire that spells it — a page summary no longer pays a reasoning pass where the endpoint thinks by default (#4107; behavior change: the summary request of `createWebFetchSummarizer` carries the dialect's own disable key where the model's wire spells `off`, and carries nothing new where it cannot; narrowing: one, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core @test)
|
|
91
|
+
|
|
92
|
+
- **Before.** The summarizer's request named no tier, so on a dialect that spells `off` (`thinking:{type:"disabled"}`, `enable_thinking:false`, `chat_template_kwargs`, `reasoning:{enabled:false}`, a declared `thinkingLevelMap.off`, the Messages budget form's `{type:"disabled"}`) the gateway's own default governed: an endpoint that thinks by default spent a reasoning pass on every fetched page. The engine's other answer-only request, the auto-mode classifier, already asked the lowest reasoning seat; this one did not.
|
|
93
|
+
- **Now.** The summarizer asks the same seat table the classifier's seat reads (`lowestReasoningSeat`): `off` where the wire spells it, written by the brain's own applier. Where the wire cannot spell it (no dialect declared and no reasoning held, a stock effort-only wire with no declared off, the adaptive Messages family, a Messages model that holds no reasoning) the request is byte-for-byte what it was, options and body: no key is guessed. A completions model that declares a dialect (`compat.thinkingFormat`) spells off whatever its `reasoning` flag says, so it now carries the disable key too. A model entry's declared `defaultThinking` is the default of a run, not of this request, so it does not govern the summary (it was not read before either). Known limits: the seat is read once from the summarizer's own model, so a `fallbacks` hop of another dialect does not get its own reading; a model is judged on its own `api`, whatever family of brain serves it.
|
|
94
|
+
- **Not changed.** The other engine-minted requests keep the tier they ask for today (the suggestions pass, the compaction summary, the host-facing `sideQuery` verb, memory consolidation, the advisor, a mod's `model.complete`): the first three are readings this change pins, none is a ruling it makes. A host's own request (a session title, a recap) is the host's to seat; `lowestReasoningSeat` and `applyReasoningToRequest` are its one table.
|
|
95
|
+
- **Who is hurt (narrowing, open to veto @server @cli @client-core).** A deployment that declared a thinking-capable dialect on the model its WebFetch summarizer runs on and relied on the endpoint's default thinking for page summaries: they are now answered without a thinking pass. There is no knob to put it back on this verb; the summarizer's model is the deployment's to choose (a different summarize-role model, or a custom `WebFetchConfig.summarize`).
|
|
96
|
+
- **Rules.** Readings of the answer-only tier: 2 → 1 for engine helpers that ask for no tier of their own (the classifier's seat and the summarizer's read `lowestReasoningSeat`). Special cases: 0.
|
|
97
|
+
- **Pins.** `test/backlog4107-answer-only-summary-tier.test.ts`: W1 W1b W2 W3 W5 D1 red on the base (no disable key on the body, no tier at the Brain seam, a declared run default does not govern the summary); W4 D2 are controls, green on both (the unspelled wires carry what they carried, whole body); R1 R2 R3 are readings of the aux requests this change leaves as they are, green on both.
|
|
98
|
+
|
|
99
|
+
### A rule's `paths:` is read in every form YAML gives a glob list, and a `paths:` that reads as no limit is said once — a list flush left (`- src/**` at column 0), a list with a blank line or a `#` comment line inside the block, a block scalar (`paths: |`) and a plain value indented under the key limit the rule to their paths where each read as no limit and showed the rule on every request (#4055, #4189; behavior change: narrowing — a conditional rule written in those forms is shown only for a matching access, as the reference loads it, open to veto @cli @server; widening: one, below — a `paths:` value that is a plain null, boolean or number, a flow list with no string item, or a plain text holding `: ` now reads as no limit; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @server @test)
|
|
100
|
+
|
|
101
|
+
- **Rule.** One function, `pathsValueOf` in `src/core/runner/nested-memory-content.ts`, reads "the value of a rule's frontmatter `paths:`" for every form, as the strings a YAML reader gives it (the reference reads the value through one: a list gives each item, a string gives the globs between its top-level commas, anything else gives none). A block list is every `-` line under the key, flush left or indented, with the blank lines and the `#` comment lines between them skipped; the block ends at the first line that is none of those. A block scalar (`paths: |`, `>`, with its chomping and indent marks) is the text of the more-indented lines under it. A plain value — on the key's line or on the more-indented lines under it — is those lines folded to one text, and a plain text that holds `: ` is a mapping's, not a value's. The reading is then the one the list items get: the commas split it, a trailing `/**` goes, the quotes come off, a plain null, boolean or number value (`~`, `null`, `true`, `12`) is no glob. The block's lines are read without their CR (a file written with CRLF ends reads as its LF twin; the body's bytes are not touched). A `paths:` that reads as no limit — no value, `[]`, `""`, nothing but comments, items none of which is a string, only `**`, a block scalar of nothing, a mapping under the key — leaves the rule applying to every file, which is the reference's delivery, and the lane says so to the operator once per session and file: the host's `onError` sink (phase `degraded`, classification `attachment-read`), one sentence that names the file and the way to limit it (list the globs, or remove the key if every file is meant). No notice code, no wire key. The sentence is one the reference does not say (registered as D-4055-a, ruled C-R1409, open to veto).
|
|
102
|
+
- **Said once — one sentence more than the reference says (open to veto @cli @server; the lightest form of "says more than the reference", the family of the no-limit sentence).** A `paths:` with a comma-separated piece written over several lines — globs on lines of their own with no comma between them, in a block scalar (`|`, `>`) or in a plain value under the key — reads as ONE glob that matches no file, so the rule is never loaded; the reference does the same without a word (the reading of its code, not measured on a binary). The lane says so on the seat of the no-limit sentence (the host's `onError`, phase `degraded`, classification `attachment-read`, once per session and file): one sentence that names the file and the way out (list the globs, or write them on one line separated by commas). The load is unchanged (the rule is still not loaded); no notice code, no wire key; `InstructionText.spansLines` is internal to the lane (not a root export). A `paths:` whose globs are separated by commas (one to a line or not), a list, a flow list or one glob says nothing.
|
|
103
|
+
- **Before.** The list reader took indented items only and stopped at the first blank or comment line, so `paths:` over a flush-left `- src/**` read as no globs, which the lane reads as "every file": a rule written for `src/` was shown on every request. A key that read as empty said nothing. A block scalar (`paths: |` over `src/**`) and a plain value indented under the key (no `-`) were read as "a form the reader does not read": no glob, so the rule was shown on every request too (measured on a real Runner: carried by the session's first request, and by every later one whatever was read) where the reference loads such a rule only after a Read of a matching file (measured on its real binary for the single-glob forms).
|
|
104
|
+
- **Widening.** A `paths:` on the key's line whose value is a plain null, boolean or number (`null`, `~`, `true`, `12`), a flow list with no string item (`[1,2]`) or a plain text holding `: ` (`src/a: b`) was read as that text taken for globs (`null`, `1` and `2`, `src/a: b`), which name no file of a real tree, so the rule was loaded only for an access to such a name; it now reads as no limit, so the rule is shown on every request and the no-limit sentence is said once — the reference's reading (a YAML reader gives no string there). Measured on both trees through `processInstructionText` (base 7.67.1: `globs`; head: `noLimit`); a quoted `"src/a: b"` is one glob on both. **Who notices.** An author who wrote such a value; the rule now reaches every request. **Compensation.** Write the globs as strings, or drop the key.
|
|
105
|
+
- **Not changed.** What the model is shown of a rule that applies; the readings of the indented list, the flow list of strings, a string scalar and the comma list; a rule with no `paths:` key; a named member (`CLAUDE.md`) that carries the key (the key limits rules, so it gets no sentence).
|
|
106
|
+
- **Who is hurt (narrowing, open to veto @cli @server).** An author who wrote the list flush left, or with a blank or comment line in it, or as a block scalar, or as a plain value under the key, meant to limit the rule, and it is now limited. A rule that relied on being shown everywhere through that misreading stops being shown for accesses outside its paths. **Compensation.** Drop the key, or list `**`.
|
|
107
|
+
- **Known limits.** An inline ` # comment` after a value is not stripped (the glob then holds the comment and matches nothing); a flow mapping (`paths: {}`) reads as one glob. A text of several lines (a block scalar or a plain value over two lines, without commas) is one glob (a literal block scalar keeps its line breaks in it, a folded one and a plain value a space), which matches no file — the reading of the reference's code, not measured on a binary — and the lane says so to the operator once (the said-once bullet above); a multi-line flow list (`paths: [a,` over the next line) is not read. This reader is not a YAML engine; the reference's is. A rule file that an `@` include of another file also names is delivered through the include (no `paths:` read as a rule there), so it gets no sentence.
|
|
108
|
+
- **Rules.** One reading of "the value of `paths:`" (`pathsValueOf`) for the list, flow list, scalar, block scalar and plain value; the flush-left and gap forms joined the indented one in the same block loop. Special cases: 0 (the mapping guard is the one rule that a plain text holding `: ` is no value).
|
|
109
|
+
- **Pins.** `test/backlog4189-paths-yaml-forms.test.ts`: the five forms through a real Runner (a rule under `.claude/rules`, the session's first request, the request after a Read under `src/`, and after a Read under `docs/`) — forms 4 and 5 are red on the base, forms 1, 2 and 3 are green on both; N1 N2 (the same two forms in a sub-directory's own rules, globs read against that level, and in a sub-directory of the rules directory: red on the base); T1 (globs over several lines — a block scalar, a folded scalar, a plain value, a block scalar with a blank line between: never loaded, the sentence said exactly once, the no-limit sentence not said: red without the sentence) and T2 (the controls with commas, a list and one glob: loaded, nothing said: green on both); U8 (the spans-lines mark, eleven forms without it, five with it); U1–U7 (every form of the block scalar and the plain value, the readings that stay no limit, CRLF, several lines, the same files matched as the list's). `test/backlog4055-4056-nested-rules.test.ts`: P1 P2 P3 P4 P6 P7 L1 and the nine E cells (an empty read is delivered to every file and told exactly once, across loads and across a session's second lane; the nine are the key with nothing under it, `[]`, `""`, comments only, non-string items, only `**`, a mapping under the key, a block scalar of nothing, a plain number under the key) are red on the base; P5 and E-control (a readable limit says nothing; a named member's key says nothing) are green on both; L2 (a CRLF rule file through the lane, which normalises the text first) was already right on this branch before the CRLF read was added.
|
|
110
|
+
|
|
111
|
+
### A rules directory larger than one load's probes is delivered whole across loads — what the lane asked about before (a delivered member's check, a rule that showed nothing at its turn, a subdirectory's listing) waits behind the waiting ones of its level, so it no longer keeps the members not yet reached from being probed (#4056; behavior change: rules past the first loads' probes are now delivered where they never were, each through the admission it always had; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @test)
|
|
112
|
+
|
|
113
|
+
- **Rule.** The lane's law is "what is left waits for the next load": a load makes at most 32 probes, and what it did not reach stays queued. The root's level is asked on every load, so a rule or a directory it had already asked about was put back at the head of the queue ahead of the ones it had not reached. Now what was asked about before goes behind the waiting items of its own level (`asked`, `asksAgain` and `queueAgain` in `src/core/runner/nested-memory-lane.ts`, read at the two places a load queues rule work: a level's own members and rule directories, and a directory listing's result); a level with nothing waiting queues them first, as a first ask does, so the outer levels keep reading before the inner ones. The checks of delivered members go on in rotation behind the first deliveries.
|
|
114
|
+
- **Before (measured on the base).** 45 small root rule files: 14, 7, 3, then 0 delivered per load, 24 of 45 in all; 100 files: the same 24 of 100; a rules directory of 31 subdirectories holding one rule each: 14 of 31; twenty rules whose `paths:` match no access followed by ten unconditional ones: none of the ten; a rule file added next to forty such members: not delivered within twelve loads.
|
|
115
|
+
- **Now.** The same 45: 14, 15, 15, 1; 100 files: 14, 15, 15, 15, 15, 15, 11 (seven loads); the 31 subdirectories: all; the ten: all; the added file: the next load. A delivered member that changed or was removed is still found by the checks (below the probes, at the next load, as before). A root rule that was asked about and then becomes applicable is still read before the accessed directory's own file.
|
|
116
|
+
- **Cost.** The root's rules directory is listed again behind what still waits, so in a directory of more than about fifteen members a file added later is found after the waiting ones were read (three loads in a directory of a hundred).
|
|
117
|
+
- **Not changed.** Which rules are delivered and their frames; the order of a first delivery within a level and across levels; the probe bound.
|
|
118
|
+
- **Rules.** One predicate (`asksAgain`) and one placement (`queueAgain`) for the two queueing places; special cases: 0.
|
|
119
|
+
- **Pins.** `test/backlog4055-4056-nested-rules.test.ts`: R1 at 45 and at 100 members, R2, R4 and R6 are red on the base; R3 (twelve rules: a change found and a removal withdrawn by the next load) and R5 (the root's level first) are green on both.
|
|
120
|
+
|
|
121
|
+
### A resume aborted after its claim names what became of its row as a word — `CheckpointError.detail.unstartedRow` — the word a failed result states as `terminal.unstartedRow`, now one declared type (#2518; behavior change: one optional member on a rejection; narrowing: none; widening: none; hard-breaking: none; closed sets: `UnstartedRowOutcome` registered (`reopened` / `declined` / `unknown`, unchanged words); exports: none; wire: `CheckpointError.detail.unstartedRow?`) (@server @client-core)
|
|
122
|
+
|
|
123
|
+
- **Rule.** What became of the checkpoint row of a resumed leg that ended before its approved action started is one closed word, declared once (`UnstartedRowOutcome`, `src/core/terminal-cause.ts`), and every exit states it as a member; the sentence is derived from the word. A failed result carries it as `terminal.unstartedRow` (the thrown-prepare and returned-leg exits, since 7.61.0, #3163); a resume aborted after its claim and before its leg started — which rejects `resume()` / `resumeStream()` instead of minting a result — now carries it as `detail.unstartedRow`: `"reopened"` on `checkpoint.resume_aborted` (the row is pending again: the same token resumes it), `"declined"` or `"unknown"` on `checkpoint.reopen_failed` (the store refused — read the row; the reopen threw and the row does not show it landed — its state is unprovable).
|
|
124
|
+
- **Before.** On that exit `checkpoint.reopen_failed` covered both `declined` and `unknown`, and only the sentence told them apart (the `resume()` docs said so); `checkpoint.resume_aborted` covered a reversed claim and a row that stays consumed alike. The three words were spelled twice, in `TerminalCause` and in the engine's own `UnstartedRow`.
|
|
125
|
+
- **Now.** Branch on the word. Without it `checkpoint.resume_aborted` does not say which state the row is in: an abort before the claim leaves it pending and untouched, an abort that found the row's park already claimed leaves it consumed — read the row (the two arms share the code and carry no member; not changed here). `UnstartedRow` is derived from the one type (a table of the facts each word carries, fenced both ways); `CheckpointError.detail.unstartedRow`, `TerminalCause.failed.unstartedRow` and `amendTerminal` read the same type.
|
|
126
|
+
- **Not changed.** The codes on every exit; `terminal.unstartedRow` on a result; the compensation (`compensateUnstarted`). The sentences are not changed by this section; #4217 (this release, below) rewrites the `declined` and `unknown` ones in words only.
|
|
127
|
+
- **Rules.** Retired: reading the row's outcome off the abort exit's sentence; the second spelling of the three words. Added: none (the member sits beside `detail.settler`, the same precedent).
|
|
128
|
+
- **Pins.** `test/backlog2518-resume-row-result.test.ts` A — A1 reopened · A2 declined (the declined sentence too) · A3 unknown (the same code; the word tells it from declined) · A4 control: an abort before the claim carries no word, the row untouched · A5 the two members are one type.
|
|
129
|
+
- **Consumers.** @server `http/resume-legs.ts` maps the abort exit's `checkpoint.reopen_failed` to one terminal posture for both forms (its own comment names the two): read `detail.unstartedRow` to tell `declined` (read the row) from `unknown` (re-read before deciding; never issue a fresh approval on the assumption the old one is dead). @client-core: a renderer that read the sentence reads the word.
|
|
130
|
+
|
|
131
|
+
### A checkpoint store's `recordExecutionOutcome` verb is the contract — the `execution = { outcome: true }` declaration retires (#2517; behavior change: the store seat reads the verb's presence; narrowing: one, below, open to veto @server; widening: one, below, open to veto @server; hard-breaking: one — `execution` removed from `CheckpointStore`, `InMemoryCheckpointStore` and `FileCheckpointStore`; closed sets: none; exports: none; wire: `CheckpointStore.execution`, `InMemoryCheckpointStore.execution`, `FileCheckpointStore.execution` removed)
|
|
132
|
+
|
|
133
|
+
- **Rule.** A store's required members are its verbs, and the store seat reads each verb's presence — one table, `REQUIRED_STORE_MEMBERS`, two rows: `recordExecutionOutcome` and `reopen` (the same shape the `reopen` verb took in 7.57.0, #1914, when the `redecision` declaration retired). `checkpointStoreIncomplete` names each missing verb.
|
|
134
|
+
- **Hard-breaking (three sentences).** *What changed:* `CheckpointStore.execution` (the declaration `execution.outcome`) is removed from the store contract and read nowhere; the seat reads `typeof store.recordExecutionOutcome === "function"`. *Who is hurt:* a TypeScript reader of `store.execution` (the in-tree `InMemoryCheckpointStore` and `FileCheckpointStore` lose the field too), and an object literal typed `CheckpointStore` that declares `execution` (an excess-property error) — @server `test/s1-smoke.test.ts` (its store literal); a class that declares the field (server's `src/plugins/local-checkpoint-store.ts`, `src/stores-sql/checkpoint-store-sql.ts`) still compiles and carries a dead field. *How to migrate:* delete the field and anything that reads it; a store wrapper forwards `recordExecutionOutcome` (it already must).
|
|
135
|
+
- **Widening (open to veto @server).** A store carrying every verb and declaring nothing was refused at its seat (`config.invalid_checkpoint_store`); it is accepted.
|
|
136
|
+
- **Narrowing (open to veto @server).** A store declaring `execution = { outcome: true }` without the verb was accepted at its seat and failed at its first record; it is refused at its seat, naming the verb.
|
|
137
|
+
- **Not changed.** The verb's contract (the CAS, the four words, the conflict throw, `reopen` refusing a recorded row); the refusal's code; the in-tree stores' behaviour (their field is gone). The store-contract kit's prerequisite case now names `recordExecutionOutcome` among the verbs it requires; its separate declaration case is folded into it.
|
|
138
|
+
- **Rules.** Retired: the declaration (a second representation of a required verb, which a proxy hiding the verb would still have inherited). Added: none.
|
|
139
|
+
- **Pins.** `test/backlog2518-resume-row-result.test.ts` B — B1 a store with every verb and no declaration parks, resumes and has its execution recorded · B2 a store declaring the field without the verb is refused at its seat naming the verb. Re-pinned onto the verb: `backlog602-execution-outcome` ⑦, `backlog1914-claim-reversible` S1 / S4, `backlog3688-compact-session-task-store` ④, `design393-s3-resume-lanes`; hand-listed wrappers forward the verb instead of the field (`backlog570`, `backlog1079`, `backlog1214`, `backlog2300-parked-row-mount-site`, `backlog3608`) or drop the field (`agent-park-jian4-durableapproval-forwarding`, `backlog3212`, `design252-park-selfcheck`, `design384-slice2-park-fence`, `park-mint-backend-fidelity`, `wiring-manifest`).
|
|
140
|
+
- **Consumers.** @server: delete the two `execution` fields and the literal's; nothing replaces them. @dgx: a deployment store drops the field. @cli @client-core: no store implementations.
|
|
141
|
+
|
|
142
|
+
### A resume's screen refuses three value-only arms before its claim — the posture word, per-task agents with no delegation tool, a `modelRole` outside the vocabulary — so the row stays pending and untouched (#2519; behavior change: three refusals move before the claim; narrowing: none; widening: none; hard-breaking: none; closed sets: `CHECKPOINT_CONFIG_REFUSAL_CODES` +2 (`config.interaction_posture`, `config.agents.no_delegation_tool`); exports: none; wire: two words added to the screen's refusal set) (@server @client-core)
|
|
143
|
+
|
|
144
|
+
- **Rule.** A door the configuration alone decides runs before a resume's claim: the value doors are one function with two callers (`configValueDoors` — prepare's config-door phase, and a resume's screen through `checkpoint.config_refused`), and the value arms of three doors whose other arms need the leg now sit there: the interaction posture's word outside `"interactive"` / `"headless"` and `"interactive"` beside `interactiveTools: false` (`config.interaction_posture`), the requested role's word outside the vocabulary (`config.unknown_role`, `detail.reason: "unknown_role_word"`; `screenModelRole`, the one judgment the model resolution's chain head also reads), and `TaskSpec.agents` beside a roster that mounts no delegation tool (`config.agents.no_delegation_tool`; the screen reads the roster through the receipt the tool screen uses — no caller code runs).
|
|
145
|
+
- **Before.** These refused after the claim (KL-1950-a), so the row was consumed and given back only by a `reopen` that landed (on a store whose `reopen` declines, consumed for good); `judgeResumeConfig` answered `ok` for them.
|
|
146
|
+
- **Now.** `checkpoint.config_refused` before the claim with `detail.code` the fresh run's word and `detail.field` the input that carries it (`taskConfig`, or `deps` for the deployment's posture word); the row stays `pending` and byte-identical and the corrected resume redeems the same token. `judgeResumeConfig` answers the same. A posture refusal on the screen is the resume's rejection and is not also sent to `onError` (the config question tells no sink); a fresh run still discloses it on the config lane. An entry whose `withAgents` only the caller's code can answer for (an accessor, a Proxy, an entry of another prototype) is left to the leg's own read after the claim (the KL-469-i class), and so is an `agents` value that is not a list. Whether the posture's human is reachable, the model the role resolves to and the model gate's deployment-seat shape still refuse after the claim (KL-1950-a, rewritten).
|
|
147
|
+
- **Refusal order (open to veto @server).** The posture and role words are now judged with the value doors, before the approval seat's read (`resolveAskSeat`) and the per-task agents' rebuild, on a fresh run too: a config with two faults — a bad posture word beside an `onAsk` that is no approver, say — is now refused `config.interaction_posture` (disclosed on the config lane) where it was refused `config.interaction_wiring`. A config with one fault is refused as before.
|
|
148
|
+
- **Rules.** Retired: three arms from the after-the-claim list; the posture door's second resolution of the word (prepare reads the value door's); the static wiring manifest's own copy of the word judgment (`describeStaticWiring` and the door both read `screenInteractionPostureWord`, so a fresh run's refusal now carries the manifest's sentence — "… never folded to either posture (or to absent)."). Added: none (one door function and one refusal constructor each called by both readers).
|
|
149
|
+
- **Pins.** `test/backlog2518-resume-row-result.test.ts` C — five shapes before the claim on a store whose `reopen` declines (the config's posture word, the deployment's, interactive beside `interactiveTools: false`, agents with no delegation tool, a `modelRole` outside the vocabulary), each with the row untouched and the corrected resume completing · an accessor `withAgents` is not read before the claim · the config question answers each and tells no sink · control: a fresh run's posture refusal is still disclosed.
|
|
150
|
+
- **Consumers.** @server @client-core: a table over `CheckpointConfigRefusalCode` places two words; a resume of a config with one of these arms now rejects before the claim (`checkpoint.config_refused`, row pending) instead of settling a failed leg whose row was reopened — server's resume lane already maps `checkpoint.config_refused`; `judgeResumeConfig` callers (S-3797) get the verdict before they write.
|
|
151
|
+
|
|
152
|
+
### One session's deletion has one reading of crash consistency and of the live-reference re-check, whatever store set the executor deletes from — the contract `retention.session_deletion_recorded_first` on `ManagedRetentionCapability.deleteExpiredSessions` (#3001; ruled C-R1336; behavior change: none, contract text only; hard-breaking: none; closed sets: none; exports: none; wire: none)
|
|
153
|
+
|
|
154
|
+
- **Rule.** Additive contract sentence: under an exclusion on the session that every write able to point into it takes too (a fork, a background agent, a parked checkpoint), the executor re-checks the live references, then runs the legs in `RETENTION_SESSION_TREE_STEPS` order under a deletion record (intent, audit row, tombstones) — no row it witnesses is gone while the record is not yet written and synced, which deletes committed in one transaction with the record meet as a record synced first does — marks the record done after the session row's leg, and replays a record not marked done from the first step at its next start, idempotently. It replaces the "inside the deletion transaction" wording (the capability, `EngineStorageBackend.retention`, INTEGRATION-CORE), which only a backend whose deletes commit together could meet. No code path changed here; the executor stays the deployment's, and the table it reads is `RETENTION_CASCADE` — its rows move in this release under #4000 (below) and the sections that section names, none under this one.
|
|
155
|
+
- **Pins.** `test/backlog3001-session-deletion-record.test.ts` — S1 the tag and its registry row · S2 the eight clauses · S3 no transaction-only wording left, the backend member links the clause · R a cascade over the file stores cut after each step and replayed from the first converges, and a replay after the last deletes nothing more (green on the base too).
|
|
156
|
+
|
|
157
|
+
### A fork whose row records no occupant of its source carries no file history — what a fork carries is the source occupant's history at its address, #3996's reading below; the source-row comparison this ticket first made did not ship (#3982; behavior change: a fork row with no recorded source occupant is no longer handed the source id's history; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server)
|
|
158
|
+
|
|
159
|
+
- **Rule.** A fork's file history comes from the session it was cut from, not from whoever holds that session's id when the fork first runs. The carry seat reads the fork row's lineage cell whole — the source id and the occupant of it the fork was cut from — as one address, `fileHistoryScopeOf(forkedFrom, forkedFromIncarnation)` (#3996, below), and calls `FileHistoryStore.adoptScope(<that address>, <the fork's scope>)`; it reads no row of the source. A fork row that names no occupant of its source has no source address: it carries nothing and tells `onError` once (`adopt_failed`, phase `"rewind"`, once per fork scope and history store object). The same seat serves the run (`prepareFileHistory`) and the out-of-run `runner.rewindFiles`.
|
|
160
|
+
- **Before (through 7.67.3).** The seat read only `forkedFrom` and carried by that id. The carry is lazy (it runs on the fork's first turn, which may be long after the fork), so an id erased and created again in between — a new occupant — had its boundaries carried into the fork, and the fork's first turn could `rewindFilesTo` one of them and write the shared workspace back to what another session recorded.
|
|
161
|
+
- **Did not ship.** This ticket's first form compared the source row's current occupant with the recorded one (`SessionStore.headOf`, `compareOccupants`) and carried nothing for a source row that was gone or a session store with no `headOf`. #3996 retired that comparison in this release, before it shipped: the address names the occupant, so no source row is read; a fork whose source row is gone (its history kept) and a fork in a store with no `headOf` are carried as in 7.67.1 (#3996's R-1 states it; its R-3 states the same for a source occupant recorded before the address epoch).
|
|
162
|
+
- **Narrowing (open to veto @server).** A fork whose row records no source occupant — one cut before rows named their occupants, or one in a store that mints none — carries nothing from its next prepare on, though its source row is alive. The fork starts with its own history and a pre-fork boundary answers `rewind_snapshot.unresolvable`; a fork that already holds history keeps it (KL-3982-a).
|
|
163
|
+
- **Not changed.** Same-occupant forks through the Agent tool's fork route (it carries through `RunInternals.fileHistoryLineage` and never reaches this seat); the empty-target law and the `conflict` reading; the retry of a faulted carry on the next turn; `session-checkout.ts` (it reads `forkedFrom` as a boolean only). The deletion cascade's fork legs (steps 10 and 11) read the same address function (#3996). Disclosure code and phase are existing members (`adopt_failed`, `"rewind"`); no new member, word or wire key. Stated limits: KL-3982-a (the narrowing's edges), KL-3982-d (a fork of a fork cut before its parent first prepared), KL-3992-a (a source occupant from before the address epoch is carried at the bare-id scope) and KL-3992-b (an empty source carries silently).
|
|
164
|
+
- **Rules.** Retired: the carry seat's id-only reading of the lineage. Added: none — the carry reads the one address function every history seat reads (`fileHistoryScopeOf`).
|
|
165
|
+
- **Pins.** `test/backlog3982-fork-lineage-occupant-file-history.test.ts` (re-pinned to the address by #3996) — C1 premise (the fork row records the source's occupant; the re-created id holds another) · C2 the host erased the source's history with its row: the fork restores nothing and the carry of the erased address is silent · C3 the history store kept the old occupant's graph: the fork carries exactly that occupant's history · C4 / C4b controls (the same occupant still carries, on the run seat and on `rewindFiles`) · C5 a real file row from before the lineage token carries nothing, told once, though its source row is alive · C6 an id that changed hands says nothing at the seat (no row of the source is read) · C7 a source row deleted after the fork, its history kept, is still carried · C8 `rewindFiles` never reaches the newcomer's boundary · C9 the seat's arms read directly (a recorded occupant carries from its address; none ⇒ one line; a refusal settles; a store failure is said once and retried; the memo is the fork's own scope; a non-fork makes no carry). `test/backlog1050-file-history-capture-off.test.ts` names the occupant on its stub fork.
|
|
166
|
+
|
|
167
|
+
### A user row read back from the session store says what it is — the word its `message_committed` said, `UserMessage.origin` (+ `UserMessage.inputId` on a person's row), one reader `userRowSourceOf` (root export) for the frame and the stored row (#1530; behavior change: two additive members on every new user row; a compaction's re-read or an instruction file's frame now commits as `queued_note`, below; narrowing: that one, open to veto @server; widening: none; hard-breaking: none; closed sets: none (`UserRowOrigin` unchanged, now a root value); exports: +5 `USER_ROW_ORIGINS`, `isUserRowOrigin`, `userRowSourceOf`, types `UserRowOrigin`, `UserRowSource`; wire: `UserMessage.origin?`, `UserMessage.inputId?`, additive) (@server @client-core @test @web-admin)
|
|
168
|
+
|
|
169
|
+
- **Before.** A reader of the store alone (a list's preview, a session's title, the "last thing the person said") could not tell a person's row from the engine's. Measured on a file-backed store: the objective carries no member of its own; a person's live steer, a steer whose sender wrote it (`source: "system"`) and both interruption markers carry exactly `provenance: "engine-note"`; task notifications on all three roads (`notify` live, parked then drained at the next run's open, a background command's completion in the same process) `provenance` + `engineNoteKey`; engine reminders `engineMinted`. `provenance` / `engineMinted` say how derived views present a row, not who wrote it. The commit frame knew (`message_committed.origin`, 7.67.0) from two in-process WeakMap sidecars the store never saw.
|
|
170
|
+
- **Now.** Every user row carries the word as a member, written once at its first commit point: the harness's person-input mint states `origin: "human_input"` and, when the row is the one input a `human_input` frame accounts for, its `inputId` (`personInputRow`, a hollow key refused there; a resume's continuation that merges parked steers mints its row with no `inputId`), the loop's and the run-level interruption markers state `interruption_marker`, and the leg's admission seat (`RowSeat.admitRows`, the one seat every row passes) stamps every other user row — `attachment` for a frame of an attachment a person declared, else `queued_note` (`stampUserRowOrigin`). The store keeps the members. `userRowSourceOf(message)` reads them — `{ origin, inputId?, noteKey? }` (`inputId` from the member on a `human_input` row, from the `attachment` record on an attachment frame; `noteKey` from `engineNoteKey`), the same object the row's commit frame states, which now reads the row. A row written before 7.68.0 reads `undefined`: unknown, never a word inferred from `provenance`, `engineMinted` or `engineNoteKey`. Never sent to a provider (`stripEngineMetadata` strips both); the session-import door keeps an `origin` of the set and an `inputId` beside `human_input`, drops anything else.
|
|
171
|
+
- **Narrowing (open to veto @server).** An attachment frame whose record names an engine `origin` (a landed compaction's re-read, an instruction file) belongs to no declared input: its `message_committed` said `origin: "attachment"` under the restore's (or the load's) id, which no `human_input` frame carries; it now says `queued_note` with no key, and so does its stored row. A consumer that counted those rows among a person's attachments stops doing so.
|
|
172
|
+
- **Not changed.** `USER_ROW_ORIGINS` (four words) and the judge; what every other row's commit frame says; `provenance` / `engineMinted` and every projection that reads them; the `session.append` station's `door` / `origin` (its own CC-anchored vocabulary, read off the row as before). Text a host sends as an objective or a steer — `source: "system"` included — is still an input the `human_input` arm accounts for: its row is `human_input`, and the `source` it was sent under is on its `human_input` frame, joined by the `inputId` the row now keeps. A host's own text belongs in `TaskStream.notify()` (`queued_note`) or a `wake` (#2433).
|
|
173
|
+
- **Rules.** −2 sidecars (the person-input WeakMap and the marker WeakSet retired; the mints write the row); the reader's inputs 4 → 1 (the row; the fallback "everything else is a note" retired — an unstamped row is no source); +1 stamp at the one seat. One reading of "what is this row" for the live frame and the store.
|
|
174
|
+
- **Pins.** `test/backlog1530-row-origin-on-store.test.ts` — ① objective / live steer / `source: "system"` steer read `human_input` under their keys, a `notify(next)` row `queued_note` under its frame key, and the brain never sees the members · ② the turn-scoped and run-level markers read `interruption_marker` · ③ the parked-notification road and a same-process background completion read `queued_note` · ④ final verification, boundary bundles (output style + changed files), a compaction's re-read and the git re-assert read `queued_note`, the objective's attachment frames `attachment` under its key, the compaction's product is not a user row, and the limit-approach sentence is never a stored row · ⑤ a wake merged with a parked steer reads one `human_input` row under the wake's key · ⑥ an unstamped row is no source, the seat's stamp, the import door, the strip, the root face, the retired sidecars. Every run checks row by row that the commit frame's source equals the stored row's. Re-pinned: `test/backlog3617-user-row-origin.test.ts` ① (the reader reads members), `test/backlog1698-one-notification-block.test.ts` (#2433 reader cell).
|
|
175
|
+
- **Consumers.** @server: a store-only reader (session list preview, title, delete confirmation) reads `userRowSourceOf(entry.message)?.origin === "human_input"` for "a person's row"; `undefined` = a pre-7.68.0 row (keep the old heuristic there only); a host-written row joins its `human_input` frame by `inputId` for its `source` where the row carries one (a resume's continuation that merges parked steers writes a `human_input` row with none). A store that rebuilds messages member by member must keep `origin` and `inputId`. @client-core / @web-admin: preview and insights read the same reader; the old "starts with `<system-reminder>`" guess retires for new rows. @test: rc2's FAIL cell for engine rows now reads the member; host-written text stays `human_input` by design (above).
|
|
176
|
+
- **Known limits.** KL-1530-a: the approving person's words that ride a resume's continuation with no parked steer or wake message are a `queued_note` (no `human_input` frame accounts for them, #3617's ruling). KL-1530-b: a row's `enginePrefixChars` / `engineSegments` still mark the engine's frames inside a person's row; the origin does not cut them.
|
|
177
|
+
|
|
178
|
+
### A verification gate's verifier leg runs on the implementation leg's checkpoint-store word — a store-less implementation (the word a fork child carries, a machine-started run) no longer gets a verifier that can park a card nobody will decide (#3995; behavior change: the verifier's store follows the caller's spec; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @test)
|
|
179
|
+
|
|
180
|
+
- **Rule.** `TaskSpec.checkpointStore` is the one word for "who comes back to a pause on this run": absent ⇒ the deployment's store, a store ⇒ that store, `"disabled"` ⇒ nobody. Every leg a verification gate drives runs on the word the caller's spec carries; the verifier was the one leg that did not.
|
|
181
|
+
- **Before.** The verifier's spec named no store, so on a store-less implementation it fell back to `RunnerDeps.checkpointStore`: an irreversible or egress tool the verifier called suspended it into a card in the deployment's store — a pending row nobody will decide and a pinned session — although the spec's own comment said it could never become a durable checkpoint. An implementation on its own store had its verifier's card land in the deployment's store instead of its own. The fix leg already rode the caller's spec whole.
|
|
182
|
+
- **Now.** The verifier's spec carries the caller's `checkpointStore` when the caller wrote one (absent adds no key). `"disabled"`: the verifier mints no card — a gated call takes the store-less terminal (refused fail-closed, as on any store-less run) and the verifier still answers; its own store: the card lands there; absent: the deployment's store, as before. The comment now says what is true.
|
|
183
|
+
- **Narrowing (open to veto @server).** A store-less implementation's verifier can no longer suspend on an irreversible or egress tool: that call is refused, so a verdict that needed it reads the refusal instead of waiting for an approval. A verifier card is therefore only ever minted for a run that has a store to hold it.
|
|
184
|
+
- **Not changed.** An implementation on no word (the verifier parks in the deployment's store, its `verifierPause` hand-back and `Runner.voidVerifierPause` verb as before); the verifier's headless seats (it is never given `durableApproval`); the fix leg; the gate's budgets and verdicts.
|
|
185
|
+
- **Rules.** Retired: the verifier leg's silent fall-back to the deployment's store and the false sentence that described it. Added: none — one spread of the caller's word beside the seats the leg already copies from the caller's spec. Special cases: none (the same word reads the same on the implementation, verifier and fix legs).
|
|
186
|
+
- **Pins.** `test/backlog3995-verifier-store-word.test.ts` — G1 a `"disabled"` implementation has a verifier with no card (through `runWithVerification` and `verifyCompleted`; the probe never ran) · G2 an implementation on no word parks its verifier in the deployment's store (control) · G3 an implementation on its own store has its verifier's card in that store and none in the deployment's · G4 the implementation, each verifier round and the fix leg are dispatched with the same word, for the absent word, `"disabled"` and a store.
|
|
187
|
+
- **Consumers.** @server: a store-less run's verification leaves no `verify` card (compositionLeg entry `verify`) in the pending table; nothing to change where every verified run has a store. @test: a verified run on a store-less spec whose verifier reaches an irreversible tool reads a refusal in the verifier's transcript and no pending checkpoint.
|
|
188
|
+
|
|
189
|
+
### An `unauthorized` failure's error text is the engine's one clause on every record face — the status record, the manifest row, the operator warning, a refresh's and a resource listing's per-server error no longer carry the endpoint's response body (#3997; behavior change: five faces read one engine clause where they read the remote's words; narrowing: one, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core @test)
|
|
190
|
+
|
|
191
|
+
- **Rule.** An `unauthorized` failure is the user's step, not a story the remote tells: every face that renders its error text says the one clause the call-time line carries — "The user needs to authorize this server [for the access this request needs] where the host shows the server's authorization prompt" — and never the endpoint's response body. The body is the remote's own words and may name an address (or give an instruction) that no engine sentence of the class carries; the refusal itself is the record's `errorCode` + `authorization.reason`.
|
|
192
|
+
- **Before.** The sentences about an authorization failure named no address (the arm, the call line, the notice), but the `error` text of a dial's status, the manifest row's and the operator warning's, of a refresh and of a resource listing was the remote's body as it arrived (credentials redacted, neutralized, bounded): a 401 whose body said "sign in at https://…" put that address on the status, in the manifest, on the refresh result and in the text `ListMcpResourcesTool` returned to the model, and on a refresh's partial listing as well.
|
|
193
|
+
- **Now.** The one mint of a remote-authored failure text (`remoteErrorText`) answers the class's clause when the thrown value reads `unauthorized` (a site-free reading of the classifier's rows: `authorizationOwed`), so every face that reads through it — the dial's status and manifest row, the operator warning and its `cause`, the refresh result and the partial listing's `listingIncomplete.error`, the resource listing's text and `details.errors`, a continuation page's failure — says it. The one in-place rewrite seat of a request's error object reads the same class, so the thrown request error's `cause` (the refusal it wraps) and the raw rethrow of an aborted race carry the clause as well. The call line's own clause and the record faces' clause are one function (`userMustAuthorize`).
|
|
194
|
+
- **Narrowing (open to veto @server @cli @client-core).** A reader that took the endpoint's wording or an address out of `McpServerStatus.error`, the manifest row's `error` or a refresh / listing error for a 401 or a 403-scope refusal now finds the clause; the class and the reason are on the same record (`errorCode: "unauthorized"`, `authorization.reason`) and the address stays on the structural `authorizeUrl`. Every other failure kind keeps the remote's words. A host serializer that walks a refused call's `cause` reads the clause there too, where it read the body.
|
|
195
|
+
- **Not changed.** The call-time line, the arm and the notice; the structural keys and their trust; the card (`details.authorization` is `{ reason }`); a non-authorization failure's text; the announce-once identity of a needs-authorization entry (the reason and a digest of the seat's address, not the text). A refusal wrapped in another error (the classifier reads the thrown value itself, not a cause chain) is `unknown` on the record and keeps its words, as before.
|
|
196
|
+
- **Rules.** Retired: KL-3984-a (the remote's body on the faces), and the second reading of an `unauthorized` failure's text (five faces each reading the body). Added: one site-free reader (`authorizationOwed`) and one exported clause (`userMustAuthorize`, which the call line's own sentence is now built from). Mint points: one (`remoteFailureText`, read by the status / listing mint `remoteErrorText` and by the request error's in-place seat); special cases added: none — the class is read where every face already reads the remote's words.
|
|
197
|
+
- **Pins.** `test/backlog3984-3987-authorization-address.test.ts` D1–D6 — D1 a 401 at the dial: status error, manifest row error, warning and its cause carry the clause and no address, the record keeps class and reason · D2 a refresh's failed listing and a resource listing (text and `details.errors`) · D3 a continuation page refused 401 on a refresh (`listingIncomplete.error` and `error`) · D4 the clause follows the reason (403 scope, seat holds no credential) · D5 a 500 keeps the remote's words (the law is the class's) · D6 the site-free reading agrees with the classifier at the dial over http, over stdio and on a request · D7 a refused call's thrown error and the refusal under it (its `cause`) say the clause.
|
|
198
|
+
- **Consumers.** @server: the status and manifest `error` of an `unauthorized` row is the clause; render the address from `authorization.authorizeUrl` (S-3727). @cli @client-core: a status line for such a server reads the clause. @test: a 401 body that names a URL appears on no status, manifest, refresh, listing or warning text.
|
|
199
|
+
|
|
200
|
+
### A thinking tier outside the seven is refused by name — a task's `thinking`, a role record's, an agent definition's and `resolveReasoning`'s `requested` are judged by one function, and a typo no longer runs the request on the provider's own default (#3998; behavior change: a value that passed three seats unjudged is refused at the first; narrowing: one, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core @test)
|
|
201
|
+
|
|
202
|
+
- **Rule.** The vocabulary of a thinking tier is the seven `ThinkingLevel` words. A tier is stated at four seats — a task's `thinking`, a role record's `thinking`, an agent definition's `thinking`, the resolver's `requested` — and each asks the same judge (`judgeThinkingLevel`): absent answers absent, a tier answers itself, anything else is refused (`config.invalid_argument`, `detail.field` = the seat), never read as "no tier".
|
|
203
|
+
- **Before.** A typo'd tier (`"looud"`, `"HIGH"`), `null` or a number passed the task's, the role's and the agent definition's seats unjudged (a `null` definition tier was absorbed by the child's `??` chain and ran on the parent's tier) and reached the resolver, which read it as the drop arm (`effective: "provider_default"`, `dropped: true`): the request carried no thinking parameter and a default-on gateway thought on every turn, with the run's record saying only that nothing was sent. The workflow script's `agent({ thinking })` kept a hand-written second copy of the seven words.
|
|
204
|
+
- **Now.** `runTask` is refused at prepare — `failed`, `config.invalid_argument`, naming `TaskSpec.thinking` (or `roles.<role>.thinking`) — before the model is called; `resolveReasoning` throws the same refusal naming `resolveReasoning.requested`. The drop arm is left to the cases it was written for (the model does not hold reasoning; the anthropic budget path's cap-wins skip). The script seat reads the same table (`THINKING_LEVELS` / `isThinkingLevel`); its sentence and its error class are unchanged.
|
|
205
|
+
- **Narrowing (open to veto @server @cli @client-core).** A request whose `thinking` is outside the seven — including `null` where a JSON config wrote one — used to run on the provider's default and now fails at prepare, naming the member; a role record naming such a tier now refuses the task that reads it (at the role map's door, `screenRoleMap`, so `modelFaceResolvable` answers it with the refusal too), and an agent definition naming one is refused where the engine receives it. A consumer that passed a user-typed effort word straight through sees the refusal instead of a silent default.
|
|
206
|
+
- **Not changed.** The seven tiers and `off`; an absent `thinking` (the key omitted or `undefined`); the model door's own refusal of a `defaultThinking` outside the seven; the wire appliers' reading of an out-of-contract `reasoning` is #4016's (N-2 in its section: refused by name, no request sent — reachable only by a caller that bypasses a task and hands a brain its own stream options); the clamp, the levelmap, the drop arm for a model without reasoning.
|
|
207
|
+
- **Rules.** Retired: the resolver's out-of-vocabulary drop case (the #2232 drop arm's third reason), and the workflow script's hand-kept set of the seven words. Added: one judge (`judgeThinkingLevel`), asked at four seats. Special cases added: none.
|
|
208
|
+
- **Pins.** `test/backlog3998-thinking-tier-judge.test.ts` — G1 the resolver refuses ten garbage values on eight model shapes · G2 control: the seven tiers resolve, a valid tier on a non-reasoning model is still the drop arm, the judge's three answers · G3 the task's own `thinking` is refused at prepare before the model is called; every valid tier and the absent key run · G4 a role record's tier is judged (a task's roles, a deployment's) · G5 `resolveTaskModel` judges the task's tier on both branches and a role's even when shadowed · G6 the script seat reads the same table · G7 an agent definition's tier is judged at the receive door (the per-task agents seat refuses prepare before the model is called; the delegation tool refuses at construction). Re-pinned: `test/reasoning-intensity.test.ts` (#346's resolver cells now assert the refusal; the applier half is unchanged) and `test/backlog2232-effective-reasoning-provider-default.test.ts` (the out-of-contract row and tier left the drop matrix).
|
|
209
|
+
- **Consumers.** @server @cli @client-core: a run whose `thinking` / role tier is outside the seven is `failed` with `config.invalid_argument` and the member named; validate a typed effort word at your door to say it sooner. @test: `thinking: "looud"` ⇒ a loud refusal, not a run on the provider default.
|
|
210
|
+
|
|
211
|
+
### A fork leg runs with the checkpoint store switched off — a fork child and every leg it spawns are store-less, so no pause in a fork's subtree leaves a card nobody can decide (#3989, #3981 S1; ruled C-R1334 K1 (c); behavior change: a fork's gated ask is answered by a live seat or refused fail-closed instead of parking, a fork's plan-review request is dropped, the background fork row settles by the child's own outcome; narrowing: one, below, open to veto @server; widening: none — this section's first form widened a forced gate's fork leg to its live approver, and #4111 in this release refuses that ask instead, below; hard-breaking: one, three sentences below, open to veto @server; closed sets: none; exports: none; wire: none) (@server @client-core @cli @test)
|
|
212
|
+
|
|
213
|
+
- **Rule.** A pause is for a leg someone can come back to continue. Nothing continues a fork child — revival refuses the type (`revive.fork_unsupported`) and no SendMessage tier continues one — so the fork lane builds its child spec at one mint (`forkChildSpec`, `src/agents/subagent.ts`) that carries the existing per-run off switch `TaskSpec.checkpointStore: "disabled"`, whatever seat the parent resolved. Every pause family needs a store to commit (approval, content question, plan review, resource, platform), so none can arise on the leg, and each ask takes the terminal every store-less leg already takes: a live seat answers it in-stream, otherwise it is refused fail-closed with `delegation.ask_unresolvable` (`parkLaneExisted: false`). The word presses on the subtree (it is the seat a leg hands its children), so every leg a fork spawns is store-less too. It is a run's word, not a session's: a descendant that another run revives later (SendMessage's durable revival from the root session) is rebuilt under that run's seat, as every revival is, and a pause there is the plain lane's park, which that run can decide.
|
|
214
|
+
- **Before.** A fork child (synchronous or background) at a gate committed a pending checkpoint that the deployment could list and nobody could decide: the background row settled `failed` with `error: "suspended"` and a `resumable: false` frame while the parked-row lane answered `not_parked`; the synchronous parent read "This delegation is PARKED at an approval gate … Only the deployment's approval channel can release it". The same held for a plan-review request (a `plan_review` card), for a content question under a durable question gate, and for the approval asks of a leg the fork delegated to. Under `forceDurableGate` the fork parked before a live approver was asked (asked 0 times), so the action never ran.
|
|
215
|
+
- **Now.** No card in any fork cell. The background fork row and its frame settle `completed`; the synchronous fork card reads `status: completed` with the refused action in the child's result; a leg the fork delegates to is refused the same way; under `forceDurableGate` the leg announces `config.durable_gate_unavailable` once with `cause: "leg_not_resumable"` (the leg is read before the store, #4111: a fork leg is a leg nobody resumes, whatever store it was handed) and its ask is refused without consulting the approver; an ancestor policy layer under a durable mandate refuses the fork's ask `mandate_unreconstructible` (`parkLaneExisted: false`, was `true`).
|
|
216
|
+
- **The notice names both setters of the word.** `config.durable_gate_unavailable`'s `task_store_disabled` sentence named only the host ("this task disabled the checkpoint store … Run this task with its store"), which a fork child cannot act on. It now reads "this task runs with the checkpoint store switched off (checkpointStore: "disabled" — set by the host on the task, or by the engine on a fork child and every leg under it, since nothing resumes a fork)", and its way out "Run the task with its store (a fork child cannot have one: delegate work that must wait for a person without subagent_type "fork")" — one sentence for every leg, no fork branch. `detail` is unchanged.
|
|
217
|
+
- **Hard-breaking — one, three sentences (open to veto @server).** *What changed:* a fork child (synchronous or background) and every leg it spawns run store-less — they mint no checkpoint, a synchronous fork card never reads `status: suspended` / PARKED, a background fork row never settles `failed · suspended · resumable: false` for a pause, and a fork's plan-review request is dropped. *Who is affected:* @server's fork-card pins (`test/fork-park-card-s3460-e2e.test.ts` ② ③) and the S-3460 acceptance cells ③ ④, @test's black-box cells that read a fork card or its PARKED report, a host list or display that showed fork cards (they were never decidable), and a deployment whose fork children delegated work that parked for approval. *What to do:* re-pin fork readings to the new terminal (no card; `delegation.ask_unresolvable` with `parkLaneExisted: false`; the fork settles by its own outcome); delegate work that must wait for a person without `subagent_type: "fork"` (a plain background child's card stays decidable).
|
|
218
|
+
- **No widening (this section's first form had one; #4111 in this release replaces it).** The first form let a `forceDurableGate` principal's fork leg ask its live approver in-stream. With #4111 a fork leg is a leg nobody resumes, so under the forced gate its ask is refused without consulting the approver (`src/core/runner/prepare-ask-lane.ts`, the forced-gate arm; Now above; pin X1). Against 7.67.1 the gated action still does not run — it parked a card nobody could decide then, it is refused with the way out now. An ancestor policy layer's durable mandate is refused `mandate_unreconstructible`, as before.
|
|
219
|
+
- **Narrowing (open to veto @server).** A leg a fork spawns — against the fork directive: a plain synchronous or background child, a workflow's agents, an observer sidecar on such a leg — no longer parks for approval: a live seat answers, otherwise it is refused (KL-3989-b). Before, its card was decidable through the plain lanes.
|
|
220
|
+
- **Not changed.** The plain background lanes and their parks (this section does not touch the synchronous lane; #4111 in this release makes a synchronous child park nothing, below); content asks with no question gate (the in-tool "No human is available" sentence, fork included); revival's fork refusal; closed sets, exports and wire (`gate:wire` reads no new key). The refusal sentence of a content question on a store-less leg ("could not be parked … no live face answered it", worded so by #4012 in this release, above) is not changed here.
|
|
221
|
+
- **Rules.** Retired: the synchronous fork lane's durable-pause report arm (unreachable by construction) and one of the two hand-kept fork child-spec literals (one mint now; the off switch written once). Added: none — one existing field value at one mint. "Fork children are not forwarded the durable opt-in" is now a structural fact (no store to forward to), not a comment rule.
|
|
222
|
+
- **Known limits.** KL-3989-a: a card a fork minted before this release stays pending until its TTL — no engine lane decides it (revival refuses the type, the parked-row lane answers `not_parked`; the row carries no fork fact; no startup sweep), though a host that calls `Runner.resume` on it directly still continues it. KL-3989-b: the whole fork subtree is store-less for the runs the fork starts.
|
|
223
|
+
- **Pins.** `test/backlog3981-fork-parked-card-dead-end.test.ts` (30 cells): T1 / T2 the background / synchronous fork leave no card and settle by the child's outcome · S1 / S2 under a parent store seat, no seat and a `"disabled"` seat the fork child spec carries `"disabled"` (sync / background) · R1 the fork lane names neither the pause guard nor the pause report, and writes the word once · X1 forced gate + live approver: never asked — refused unconsulted, the action does not run, no card, `leg_not_resumable` announced · X2 / X6 forced gate + unavailable (sync / background): refused, no card, both notices · X3 no mandate · X4 plan review · X5 the seat a fork hands its children · G1 a real grandchild run · X16 a content question under the durable question gate · M1 / M2 an ancestor policy layer's durable mandate (forced / the parent's own opt-in) · controls P1, S3, X7–X15. Base (stage head `30e206710`): 17 red, 13 green (the 11 controls and the two `"disabled"`-seat spec cells); 30 green here. Nine mutations, every cell red under at least one: the mint without the word (17 red), the background or the synchronous run reading the parent's seat (3 / 15), the synchronous pause arm restored (R1), the old notice sentence (X1 and the 522 cell), the children's seat reading the store (X5, G1), every lane switched off (the plain-lane controls), the content-question sentence changed (X7, X8, X10–X13), the mint dropping the seat (the `"disabled"`-seat spec cells). Re-pinned: `test/backlog522-durable-gate-loud.test.ts` (the sentence), `test/subagent-durable-pause-disclosure.test.ts` (the fork cell: no pause report, the spec carries the word).
|
|
224
|
+
- **Consumers.** @server S-3460: flip ② ③ to "no card"; acceptance ③ reads a fork's plan review / platform pause as no card and the store-less terminal, ④ reads a fork's own forced gate as a refusal, the live approver never consulted, plus one `config.durable_gate_unavailable` (only an ancestor policy's durable mandate is `mandate_unreconstructible`). @client-core: fork cards no longer appear on the approval list. @cli: a synchronous fork reports through the ordinary card. @test: the fork-card black-box cells re-read to the new terminal.
|
|
225
|
+
|
|
226
|
+
### An authorization failure names no address in any sentence — the `mcp_unavailable` frame's needs-authorization arm, the line a refused call returns and the `mcp.authorization_required` notice's message point at the host's own prompt, and the address rides only the structural `authorizeUrl` (#3984; ruled C-R1323; behavior change: three sentences no longer spell the seat's address, and the arm no longer lists it beside the server; narrowing: one, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core)
|
|
227
|
+
|
|
228
|
+
- **Rule.** The address of an authorization has one carrier on the live faces, the structural key `authorizeUrl` (the notice's `detail`, the status record, the manifest row); no sentence a person or the model reads carries it. One phrase (`WHERE_TO_AUTHORIZE`: "where the host shows the server's authorization prompt") is where every such sentence says the user authorizes; it interpolates nothing.
|
|
229
|
+
- **Before.** The seat's `authorizeUrl` was spelled into three sentences as well as carried by the key — the frame arm's server line and its closing sentence (`<server> — authorize at <url>`, "at the address listed beside it"), the refused call's line ("authorize this server at <url>") and the notice's message ("authorize it at <url>") — so a deployment that withheld or rewrote the key still sent the address in prose, and the arm and the lines read differently with and without an address.
|
|
230
|
+
- **Now.** `needsAuthorizationArm` lists the servers by name alone and closes with "tell them that it needs to be authorized where the host shows the server's authorization prompt"; `unauthorizedLine` and the notice's message say the same phrase. A seat that gave no address and one that gave one read word for word alike. The endpoint's `resource_metadata` was never rendered and still is not; a refused request's response body is not relayed by these sentences either.
|
|
231
|
+
- **Narrowing (open to veto @server @cli @client-core).** A reader that took the address from a sentence now finds none: the model and the person see a pointer to the host's prompt, and the address is `detail.authorizeUrl` on the notice and `authorization.authorizeUrl` on the status record and the manifest row, which a host renders where it renders authorization. The three consumers stated that they read the structural keys (server: the notice's `detail.authorizeUrl` is forwarded as written, the status face follows with its authorization-record work; client-core: the notice fact, `mcpLiveness.authorization.authorizeUrl` and the seat contract's outbound form; cli: the manifest row's `authorization.authorizeUrl` through `readMcpAuthorization`).
|
|
232
|
+
- **Not changed.** The structural keys and their trust (`authorizeUrl` only from the seat, `resourceMetadataUrl` only from the endpoint's challenge); the announce-once identity of a told needs-authorization entry (the reason and a digest of the seat's address, as before: `needs authorization (rejected, 8887cbbdbb175887)` for the pin's fake address); the `mcp.server_unavailable` notice (it never carried an address); the configuration refusal that names an endpoint's origin (a configuration error says which endpoint; it is not an authorization failure); the delivery tails and the closing "Do not ask the user for authorization codes, tokens, or callback URLs".
|
|
233
|
+
- **Rules.** Retired: the one interpolating function (`whereToAuthorize(authorizeUrl)`), the arm's with-address / without-address / mixed phrasings and the "no authorization address was given" fact — three phrasings become one. Added: one constant.
|
|
234
|
+
- **Disclosed, not changed.** The told identity still digests the seat's address, so a changed address judges the same server another record and tells an arm whose words no longer differ (KL-3984-b). Until the server's status-face work lands the notice's `detail.authorizeUrl` is the host's one address exit (KL-3984-c).
|
|
235
|
+
- **Pins.** `test/backlog3984-3987-authorization-address.test.ts` — A1 the call-time line for a 401 / a 403 scope / a seat refusal, address-free, the 401 body's own URL not relayed · A2 `unauthorizedLine` over every reason and both renewal readings renders nothing beyond the reason · A3 through the Runner a dial refused 401: the arm and the notice's message address-free · A4 the arm reads the same with and without an address, the bounded form included · A5 a refused call's `tool_end` output address-free · B1 / B3 the structural keys carry the address (status record, announcer, notice `detail`, manifest row) · C1 the told identity is the same digest · D1 / D2 the remote's body in the dial, refresh and listing text, measured. Baseline red on A1–A5, B2, E1, E2; B1, B3, C1, D1, D2 are controls green on both. Mutations M1–M3, M2a–M2e, M6, M7, M9 each red on the cells they target (the arm, the line's three branches and both URL members, the notice's message, the status projection, the digest, the remote-text reader). `test/backlog2295-unauthorized.test.ts` re-pinned (the sentences; the URL law read where the address now lives, on the announcer's record).
|
|
236
|
+
- **Consumers.** @server (S-3727, S-2500): render the notice's `detail.authorizeUrl`; the status face's address is yours to project. @client-core: nothing to change (the structural keys are already rendered). @cli: nothing to change (`authorization.authorizeUrl` is read through `readMcpAuthorization`).
|
|
237
|
+
|
|
238
|
+
### The tool result card, and so the session archive, carries an authorization's reason only — `details.authorization` of a refused MCP request is `{ reason }`, and the seat's `authorizeUrl` and the endpoint's `resourceMetadataUrl` are no longer written into it (#3987; ruled C-R1326; behavior change: the card of an `unauthorized` request failure loses two optional members; narrowing: none; widening: none; hard-breaking: one, three sentences below, open to veto @server @cli @client-core @sdk; closed sets: none; exports: none added or removed — `McpFailureDetails` is exported from `src/core/mcp.ts` for the type pin and is not a root name; wire: `details.authorization.authorizeUrl` and `details.authorization.resourceMetadataUrl` removed from the card of an `unauthorized` MCP request failure — a runtime member of the tool result, not a type the package root reaches, so `gate:wire` and `handoff:diff` read 0 for it and the two removals are declared here by hand) (@server @cli @client-core @sdk)
|
|
239
|
+
|
|
240
|
+
- **Rule.** The authorization fact has two live readers and the archive is not one: the notice's `detail` and the status record carry it, and the tool result card — which the run loop copies from the thrown error's `details` and the session keeps byte for byte — carries `Pick<McpAuthorization, "reason">`. A lossless archive that no deployment disposition can edit must not hold an address the engine wrote (the endpoint's `resourceMetadataUrl` is the remote's own claim, untrusted).
|
|
241
|
+
- **Before.** `details()` copied the whole `McpAuthorization` onto every `unauthorized` request failure's card: `reason`, the seat's `authorizeUrl` and the endpoint's `resourceMetadataUrl`. The card became the tool result's `details` and so a session entry; a deployment that strips the two members at its own disposition table could not reach the archived copy.
|
|
242
|
+
- **Now.** The card is `{ reason }`. One reader of the card is gone with the members: the tool-call lane's one announcement — made once its attempts are over, after the error is thrown — used to read the whole record back off the card to feed the `mcp.authorization_required` notice. It now reads the record off the refusal the thrown error carries as its `cause`, with the one classifier every face reads a refusal by (`classifyMcpFailure`), so the notice's `detail.authorizeUrl` for a refused call is unchanged (pin B2, and E2 on the notice through the Runner). The status record, the manifest row and the notice are untouched.
|
|
243
|
+
- **Hard-breaking — one, three sentences (open to veto @server @cli @client-core @sdk).** *What changed:* `details.authorization` on the tool result card of an `unauthorized` MCP request failure has `reason` only; `authorizeUrl` and `resourceMetadataUrl` are not written into it, so they are not in the archived session entry either. *Who is affected:* a reader of `details.authorization.authorizeUrl` or `.resourceMetadataUrl` on a tool result card or an archived entry (the three consumers' stated reads are the notice and the status keys, so card readers are expected to be none — inferred from their statements, not verified in their code), and a host that stripped the two members at a deployment disposition table (nothing is left to strip). *What to do:* read the address from the notice's `detail.authorizeUrl` or from the status record's / manifest row's `authorization.authorizeUrl`; drop any disposition that strips the two members from archived cards.
|
|
244
|
+
- **The earlier staging retires.** The per-deployment strip of the two members, accepted as a staged arrangement until the server's authorization-record work lands, is retired by this release: the archive no longer carries them, so the strip has nothing to act on (server: no change needed on arrival).
|
|
245
|
+
- **Not changed.** The status record (`McpServerStatus.authorization`) and the manifest row keep both members; the notice's `detail` keeps `authorizeUrl`; the card's other keys (`code`, `delivered`, `server`, `httpStatus`, the timeout bits); the thrown error's `cause` (the raw refusal, never copied onto the card or into the archive: the one extractor of a thrown value's marks reads `details` and `errorKind` only).
|
|
246
|
+
- **Rules.** Retired: the card as a carrier of the address (carriers 3 → 2) and the announcement's read-back of it (readers of the card: 1 → 0). Added: none — the announcement reads the refusal by the classifier that already reads it.
|
|
247
|
+
- **Pins.** `test/backlog3984-3987-authorization-address.test.ts` — E1 `details.authorization` is exactly `{ reason }` for a 401 with both URLs, a 403 scope with the metadata url and a seat refusal with the seat's address · E2 through the Runner the persisted session entries carry no address of a refused call, the card keeps the reason, the live notice keeps the address · E3 the declared type has no URL member (`expectTypeOf`, read by tsc: baseline red with 2 errors, branch clean, the type widened back red with 3) · B2 a refused call's announcer is handed the whole record while the card carries the reason · B1 / B3 the status record and the notice keep the address (controls). Baseline red on E1, E2, B2; mutations M4 (the card copies the record) red on E1, E2, B2, M8 (the announcement reads the card back, as it did) red on B2, E2, M5 (the notice drops the address) red on B3, E2.
|
|
248
|
+
- **Consumers.** @server (S-2500, the sync archive): the archive is address-free from this release; no disposition needed. @cli @client-core: no card read; the structural keys are as before. @sdk: nothing read from the card is known (inferred).
|
|
249
|
+
|
|
250
|
+
### A request that put no thinking parameter on the wire resolves to `effective: "provider_default"`, not `off` — the drop arm of `TaskResult.effectiveReasoning` and the `reasoning.resolved` frame says "nothing was sent, the provider's own default governs" instead of claiming a measured off (#2232; cli L-1737 / X2; behavior change: the drop arm's `effective` reads `provider_default` where it read `off`; narrowing: none; widening: one, below, open to veto @cli @client-core @server; hard-breaking: one, below, open to veto @client-core; closed sets: one new vocabulary, `ReasoningEffective` (the seven tiers + `provider_default`; read by the wire manifest); exports: none at the root — `ResolvedReasoning` keeps its name and keys, `effective` widens from `ThinkingLevel` to `ReasoningEffective`, on the frame too; wire: none beyond the one word — no key added or removed on the result or the frame) (@cli @client-core @server @test)
|
|
251
|
+
|
|
252
|
+
- **Rule.** On the drop arm `effective` says what the WIRE carried — nothing. A request that reached the wire says its tier (clamped or not). A request that put no thinking parameter on the wire — the model's entry declares no reasoning capability (`false` or absent), or the anthropic budget path's hard output cap is too small to host a thinking window — says `provider_default`: the engine does not know whether that provider reasons on a request that names nothing, and "off" is a word the wire speaks (a disable key), never one an absent key speaks. `dropped: true` and the word are one fact, minted together by ONE function (`droppedResolution`, `src/brain/reasoning.ts`) that both drop sites (the entry gate and the cap-wins skip) call.
|
|
253
|
+
- **Before.** The two drop sites each spelled `effective: "off"` beside `dropped: true`. A run on a model whose entry declares no reasoning, asked for `--effort low`, streamed reasoning frames from a default-on gateway while its receipt said `effective: "off"` (the test lane's c164g reading).
|
|
254
|
+
- **Now.** `ResolvedReasoning.effective` is `ReasoningEffective` = the seven tiers or `provider_default` (and `on`, which #4016 in this release adds, below); the word appears only on the `dropped: true` arm, and `dropped: true` only with the word. The trace frame (`reasoning.resolved`) and the result seat (`TaskResult.effectiveReasoning`) are the two reads of the one published resolution (`publishReasoningResolution`) — both carry the word, including after the first-committed-request correction. Every other arm is unchanged by this section: the binary intent echo (`graded: false`, `effective` = the request), the levelmap-null and no-effort-value echoes, the clamp, the budget gradient, and an explicit `off` request (`effective: "off"`, no `dropped`). #4016 in this release (below) then rewrites them from the keys the wire carries: a switch-only shape says `on` where it echoed the tier (its HB-1), a withheld carrier or an `off` the wire cannot spell says `provider_default` / `dropped` (HB-2), and an `off` leg carries `effectiveReasoning` and a `reasoning.resolved` frame (HB-4; `src/core/runner/run-reasoning-seat.ts` resolves every tier a leg holds).
|
|
255
|
+
- **Widening (open to veto @cli @client-core @server).** `effective` on the drop arm now takes a value outside `ThinkingLevel`; a consumer that typed the field as the seven tiers sees one more.
|
|
256
|
+
- **Hard-breaking — one, three sentences (open to veto @client-core).** *What changed:* `effective` on a `dropped: true` resolution is `"provider_default"`, where it was `"off"`. *Who is affected:* a consumer that reads `effective` against the seven tiers and refuses or discards an unknown word (the ticket names client-core's `readEffectiveReasoning` narrow read of the result's `effectiveReasoning`), a host that switches exhaustively over `ResolvedReasoning["effective"]` or the frame's `effective` (the `reasoning.resolved` row of `TraceEvent`), and one that took `effective === "off"` on a dropped resolution as "thinking is off". *What to do:* accept `provider_default` (meaning: the provider's own default governs, unknown to the engine) beside the tiers, key "thinking is off for certain" on an explicit `off` request, and read `dropped: true` where the arm matters; a shell that expands the frame whole needs no change (the word renders as it is).
|
|
257
|
+
- **Not changed.** Every other `effective` value; `dropped`'s presence rule; `graded` / `clamped` on the drop arm (`false` / `true`); `format` / `endpoint` (the family that WOULD have carried the tier); the wire (no request changes); the trace frame's version (`1`). Read the drop arm as the one that says `provider_default`: the intent-echo arms on a wire with no carrier (stock completions or the Responses lane with `supportsReasoningEffort: false`, or a `thinkingLevelMap` null entry for the tier) also put no parameter on the wire and still report the request as `effective` beside `graded: false` — this section leaves them as they were; #4016 (below, HB-2) then makes them `provider_default` / `dropped`.
|
|
258
|
+
- **Rules.** Retired: the second literal of the drop shape (the cap-wins arm's own `effective: "off"` object). Added: one type alias (`ReasoningEffective`), one function (`droppedResolution`) and one shared-keys interface (`ReasoningResolutionKeys`, a module export that is not a root export — `ReasoningResolution` and `ResolvedReasoning` declare `requested` / `graded` / `clamped` once); the word's literal is spelled once in code and a pin counts that. Mint points: one (`droppedResolution`, called from the two drop sites); special cases added: none — the word is the same on every format (a provider whose default is known to be off still says what the engine knows, which is that the provider's default governs).
|
|
259
|
+
- **Pins.** `test/backlog2232-effective-reasoning-provider-default.test.ts` — G1 every shape that never reaches the wire resolves to `provider_default` (six resolver rows + the end-to-end undeclared-reasoning-plus-low-effort run, both faces) · G2 an explicit `off` stays `off` with no `dropped` (four resolver rows + the run mints nothing) · G3 every arm that is not the drop arm keeps its tier word (eight cases, the intent echoes included) and `dropped` ⇔ `provider_default` with `off` said only of an `off` request, over fifteen model shapes × seven tiers × four request facts · G4 the trace frame and the result seat carry one word (entry drop, the real anthropic brain's first-committed-request correction, a request that reached the wire) and the literal is spelled once in code. Re-pinned (the old word → the new): `test/observability.test.ts`, `test/backlog367-reasoning-per-request-facts.test.ts`, `test/reasoning-intensity.test.ts`, `test/reasoning-levelmap-resolver.test.ts`, `test/backlog688-contract-batch.test.ts`, `test/anthropic-fable5.test.ts`.
|
|
260
|
+
- **Consumers.** @cli (L-1737): the receipt line for a dropped resolution can say "provider default" instead of "off"; the frame expands whole, so nothing else moves. @client-core: `readEffectiveReasoning` accepts `provider_default` beside the seven tiers. @server: forwards the result as it is. @test: c164g recheck — the same run's receipt reads `provider_default`, not `off`, while the stream carries reasoning frames.
|
|
261
|
+
|
|
262
|
+
### A park delivery to a workflow run that is ending answers `suspending` — the run's closing window no longer reads `not_active` — and `not_active` means only "this process has no such run" (#4017; behavior change: `redeemWorkflowParks` between a run's terminal latch and its end answers `{ kind: "suspending", ended }` where it answered `{ kind: "not_active" }`, and the `WorkflowParkDelivery` d.ts sentences say the same; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — the members of `WorkflowParkDelivery` and their shapes are unchanged) (@server @cli @client-core)
|
|
263
|
+
|
|
264
|
+
- **Rule.** One word, one fact. `live`: this process's run took the decisions. `suspending`: this process's run is ending — suspended on its parks, or finishing — and `ended` settles once the write of its terminal row has settled, after which the host acts on the record. `not_active`: this process has no run of that id and scope (it ended and is gone, it runs in another process, or it never ran here).
|
|
265
|
+
- **Before.** The run latches its terminal boundary before it writes its terminal row. A delivery in that window answered `not_active` — "no run of that id and scope is running in this process" — while the run was running here and its record still read `running`: the behaviour, the d.ts sentence and the record gave three readings of one moment, and the host had no handle to wait on before acting on the record.
|
|
266
|
+
- **Now.** The delivery reads the terminal latch beside the suspension latch and answers `suspending` with the one `ended` promise (the run's `done`, which settles after the terminal row's write and the run's own cleanup; the active mark is dropped in the same step, so a delivery after it answers `not_active`). The decisions are not applied in the window: once `ended` settles the host reads the record and acts on it (resumes the run with the decisions if the record says it suspended). A delivery that was already past its first look — it took a park off its seat and was reading the checkpoint store — when the run began ending answers the same `suspending` with the same `ended` (it answered `live` with `not_parked_here` for that decision); a decision of it applied before the run began ending is in the run's record, not in the answer.
|
|
267
|
+
- **Disclosed.** The terminal row's write is the run's own best-effort persist (a store that refuses it, or answers that the revision moved, leaves the record as it was, as for every write of the run; a run built with no run store has no row at all): `ended` settles either way, so it says the write has settled — written or refused — never that the row is terminal; the host reads the record after it. The sentence in `WorkflowParkDelivery` says so.
|
|
268
|
+
- **Not changed.** The members of `WorkflowParkDelivery` and their shapes; `live` and `not_armed`; a delivery under another scope or another run id (`not_active`); the decision's own disposition when the run is still live (`driving`, `settled`, `kept_parked`, `already_taken`, `not_parked_here`, `unreadable`).
|
|
269
|
+
- **Rules.** Retired: the closing window's own answer (one arm of the delivery's reading, and the sentence that had to explain it). Added: none.
|
|
270
|
+
- **Pins.** `test/backlog4017-redeem-closing-window.test.ts` — G1a / G1b the closing window of a completed / a failed run (terminal write withheld by the store): the delivery answers `suspending`, `ended` stays pending while the record reads `running` and settles only after the terminal row is written (the record `ended` settled on is the terminal one), and the run is then gone from the door · G2 the stop window (a leg parked, the sibling done) answers `suspending` the same way (control) · G3 an ended run, another scope in the closing window and another run id answer `not_active` (control) · G4a / G4b the d.ts sentences: `suspending` names the closing window, the terminal row and "written or refused by the store", `not_active` no longer says "is running" · G5 a store that refuses the terminal row: `ended` still settles and the record keeps reading `running` · G6 a delivery held in its checkpoint-store read while the run is cancelled (it suspends on another park) answers `suspending` with `ended`, not `live`.
|
|
271
|
+
- **Consumers.** @server (S-3469 multi-replica park redemption): a `suspending` answer now covers a run that is finishing as well as one suspended on its parks — wait on `ended`, read the run's record, then act (resume with the decisions if it suspended; nothing to resume if it completed); a delivery that answers `not_active` is a run this process does not hold. @cli @client-core: read the two sentences in `WorkflowParkDelivery`'s comment.
|
|
272
|
+
|
|
273
|
+
### The engine's question tool carries one content ask to one mode read point, whichever form delivers the question — under `dontAsk` an `AskUserQuestion` call no layer asked about is refused by the mode before its face is reached, in the durable form's sentence and record (#4026; behavior change: under `dontAsk` the question tool's live form and its face-less mount are refused instead of shown / self-answered; narrowing: one, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — the refusal's disposition is the one the durable form already carried) (@server @cli @client-core)
|
|
274
|
+
|
|
275
|
+
- **Rule.** A call to the engine's question tool is a content ask by the tool's nature (`questionToolAsk`, the one mint the durable question policy answers with). The gate's mode read point of the first adjudication (`readModeOn`) reads it in both forms: the surviving ask of a call some layer asked about (the durable form, unchanged), and a call no layer asked about on its way to the face the tool calls itself (the live form) — classified by the one class derivation over the call's facts and judged by the one `arbitrateByMode`. An `ask` verdict changes nothing; only a refusal moves the call.
|
|
276
|
+
- **Before.** The mode read the question only where a layer turned the call into an ask, and the question tool became an ask only when the deployment composed `createDurableQuestionPolicy()`. Under `dontAsk` a durable deployment's question was refused by the mode while a non-durable deployment's was put in front of a person: one mode, two answers. The reference implementation answers once — its question tool's permission check answers `ask` and its pipeline turns an ask into a deny under dontAsk.
|
|
277
|
+
- **Now.** Under `dontAsk` every call to the engine's question tool is refused before any card is shown or any row is parked: `tool_end.gate.disposition` `{ kind: "denied", deniedBy: "mode", mode: "dontAsk", askClass: "content_question", classes: ["content_question"] }` and the sentence `Permission to use AskUserQuestion has been denied because this run is in the dontAsk permission mode (this call needed approval: content_question). …` — one frame for a live face, a durable park (with or without a live face) and a run that mounted the tool with no face.
|
|
278
|
+
- **Narrowing (open to veto @server @cli @client-core).** A `dontAsk` run with a live question face (`onQuestion`) and no durable question policy no longer reaches the face; the model reads the mode's refusal. A `dontAsk` run that mounted the tool with no face (`interactiveTools: true`) no longer receives the self-answer continuation (`Proceed with your best judgment …`); it receives the same refusal.
|
|
279
|
+
- **Not changed.** `default`, `acceptEdits`, `plan`, `auto` and `bypassPermissions` hand the question on as before (the content class is bypass-immune, and no classifier, sandbox or allow rule answers it): the tool calls its live face, the durable form parks or is routed to the face. Each form's no-answer exits — the tool's self-answer continuation and posture-coded failures; the park, its re-route and the not-shown refusal. A hook's own ask under `dontAsk` (the floor arm; on the question tool the content class floors it, as before). A caller's tool named `AskUserQuestion` (an ordinary caller tool). The resumed question call (a resume does not re-run the gate). The approved-edit pass's own read point and sentence.
|
|
280
|
+
- **Rules.** Retired: the second reading of "does the mode answer a question" (by whether a policy asked). Added: none — the live form reads the existing read point, class derivation, arbitration and refusal; the first adjudication's read point became one function both forms call, and the durable policy answers with the tool's own ask instead of spelling one.
|
|
281
|
+
- **Pins.** `test/backlog4026-live-question-mode.test.ts` — G1 dontAsk × a live face, no layer asks: refused by the mode, the face never reached, the frame equal to the durable form's · G1b dontAsk × no face: the same frame, no continuation · G1c a caller's tool named `AskUserQuestion` runs under dontAsk · G2 / G2b dontAsk × the durable form with and without a live face: refused, nothing parks · G3 the five other modes × a live face: the face answers once, the call allowed, nothing released; G3d default × the durable form is routed to the face · G3u / G3x / G3f / G2p default × each no-answer exit (face reports nobody, face throws, no face, durable with no face) stays its form's own · G4a a hook's own ask on another tool is still put to a person under dontAsk; G4b a hook's ask on the question tool is refused · G5 (AST) the first adjudication's arbitration has one call site inside the read point both forms call; the live form hands it the tool's own ask; the durable policy answers with it; no module of the question's delivery path reads the mode. Baseline red on G1 / G1b / G5 ×2, the 15 controls green; green here. Mutations M1–M5 each red (the live block never runs: G1 G1b; the live form judged by the name: G1c G5; the live form's ask put on the call: G3u G3f; a judge and sentence of the live form's own: G1 G1b G5; the durable policy spelling its own ask: G5).
|
|
282
|
+
- **Consumers.** @server (the 7.150.0 disclosure pin): the `dontAsk` × non-durable question row flips from "shown" to "refused by the mode"; the two forms now carry one frame. @cli @client-core: a `dontAsk` session renders no question card; render the refusal from `gate.disposition` (`deniedBy: "mode"`, `askClass: "content_question"`). @test (black box): under `dontAsk` both forms of `AskUserQuestion` are refused with one sentence and one disposition.
|
|
283
|
+
|
|
284
|
+
### A refused parked call raises `classic.PermissionDenied` on the resumed leg — the decide's own `deny` and every re-check that vetoes an approval, with the record its `tool_end` carries (#4052; server S-3416 r1 §7; behavior change: the deny station is raised on the resumed leg's twelve refusal arms, which raised nothing; narrowing: none; widening: one (open to veto @server @cli); hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core)
|
|
285
|
+
|
|
286
|
+
- **Rule.** A parked call's resumed execution has one refusal exit, the live deny exit's twin: the record (the decided one for the decide's own refusal; one more gate record, minted as the veto of what the decide settled, for a re-check's), then the PermissionDenied station with that record, then the `tool_end` frame and the transcript row. The payload's `input` is the exit's frozen record of the decided args (the frame's `refusedInput`), its `reason` the raw refusal sentence the result wraps, beside the leg's identity and observed mode. Every PermissionDenied payload is now minted by one function (`permissionDeniedOf`) and raised through one station body (`observePermissionDenied`).
|
|
287
|
+
- **Before.** The resumed leg wrote each refusal's result and returned without raising the station, so no deny observer heard of: the decide's own refusal (a person's `deny`, the host's SLA expiry, a deployment policy's refusal), the approval object no longer matching the live call, the approved call's PreToolUse re-screen refusing (a deny, a crash, silence, a refused answer), an edit the inherited chain cannot judge, the deployment policy refusing an approver's edit, an org-governed approval on a worker with no org wiring, the org re-check, the narrowing layers, the run's skill deny rows, a standing persisted deny rule, the shell read station, and the cross-boundary exit — while each of those refusals on a live call raised it.
|
|
288
|
+
- **Widening (open to veto @server @cli).** A deny observer now hears every one of those twelve arms, once per refused resumed call; an approved call that runs raises nothing, and the park itself still raises nothing (the ask domain). The server's translation station counts more PermissionDenied events on runs that resume parked approvals, and a settings `PermissionDenied` hook fires on them.
|
|
289
|
+
- **Not changed (the reference's form).** A PreToolUse handler's own `deny` of a call no person approved (the gate's phase 1, and its look at a policy's rewrite) still raises nothing: Claude Code 2.1.316 runs its PermissionDenied hooks only for the auto-mode classifier's refusals, and a PreToolUse hook's deny reaches that point as a hook-typed decision that never runs them — the exclusion is the reference's, the wider engine trigger surface is the registered divergence. The payload's members; the refusal sentences; the records on the frame and the row; the order of the frame and the row.
|
|
290
|
+
- **Rules.** Station raise sites 3 → 4 (the gate's deny exit and its fail-closed refusal, the compliance lock, the resumed call's refusal exit); predicates added: none. Retired: the resumed leg's per-arm refusal tails (wrap, settle, append, written twelve times) and the end settler's refusing parameters; the three hand-built payload literals; the compliance seat's own copy of the station body.
|
|
291
|
+
- **Reading (no new surface).** A consumer's translation table keyed on `ClassicEventName` (root export) under an exhaustive key guard is the compile-time pairing of stations and translators: a station the engine adds is a missing key in that table at the consumer's next upgrade. No runtime list is published.
|
|
292
|
+
- **Pins.** `test/backlog4052-permission-denied-station.test.ts` — ① a person's deny raises once; the record equals the frame's and the row's execution outcome, the input the frame's `refusedInput`, the reason raw · ①b an observer that throws there changes nothing (frame, row record and run stand) · ② an allow that runs raises nothing · ③a the approved-call re-screen's deny raises with the hook's record beside the person's settlement · ③b the host floor's narrowing raises with the policy's record · ④a / ④b the live phase-1 pair (a handler's own deny raises nothing, a crashed screen raises once). `test/backlog1294-approved-call-hooks.test.ts` ⑨c re-pinned: a resumed-leg veto raises, with the frame's record.
|
|
293
|
+
- **Consumers.** @server (S-3416 r1 §7): the parked-approval exclusion your MIGRATION names is lifted; the PreToolUse handler's own deny stays excluded (the reference's form), so that sentence stays. @cli: a `PermissionDenied` hook the engine raises fires on resumed refusals. @client-core: no read.
|
|
294
|
+
|
|
295
|
+
### A parked row answers "which leg asked" the way the leg's asks do — it records the leg's task id and its `fromSubagent` bit, the summary reads them, and a resume handed no parent task reads the bit back (#4083; behavior change: `CheckpointSummary.agentId` of a workflow agent with a declared `taskId` is that id, no longer its session id; widening: one, below, open to veto @server @client-core; narrowing: one, below, open to veto @server; hard-breaking: none; closed sets: `RequirementWord` + `from_subagent`; exports: none; wire: `CheckpointState.legTaskId?`, `CheckpointState.fromSubagent?`, `CheckpointSummary.fromSubagent?` — all optional) (@server @client-core @test)
|
|
296
|
+
|
|
297
|
+
- **Rule.** One fact, one answer: the leg's identity as its asks state it — the task id they name it by and the attribution bit (`AskRequest.fromSubagent`, a nameable parent task issued the leg) — is recorded at the row's one state mint from the leg's one ask source (`CheckpointState.legTaskId`, `CheckpointState.fromSubagent`), on every row the leg parks (approval, `plan_review`, `resource_limit`). The summary reads the record: `agentId` is the ask's own predicate (`agentIdOf`) over the recorded task id and `isDelegatedChild`, `CheckpointSummary.fromSubagent` echoes the bit as the literal `true`. The bit joins the leg identity members a resume reads back when it is handed no internals (`isDelegatedChild`, `rowTaskId`, the synchronous-child block): the ask source's bit is live `parentTaskId` presence OR the row's record — a record never subtracts a live parent task.
|
|
298
|
+
- **Before.** The row recorded neither. The summary re-derived `agentId` from the registry row else the session id (`legTaskIdOf({}, …)`), so a workflow agent whose spec declared a `taskId` asked as the declared id and summarized as its session id; no summary member said `fromSubagent`, so a consumer inferred it from `isDelegatedChild` — which a directly-started workflow's agent carries without a parent task (one card, two readings). A synchronous child resumed by its host with a bare config asked without `fromSubagent` while its row still named its parent task.
|
|
299
|
+
- **Now.** R1: a directly-started workflow's agent with a declared `taskId` — ask `agentId` = summary `agentId` = the declared id, `fromSubagent` absent on both (base: summary `agentId` = the session id). R2: a workflow under a host task — ask and summary both `fromSubagent: true` (base: summary absent). R3: a background child on a lane that can park parks, is resumed with no internals, parks again (a synchronous child parks nothing since #4111 in this release; the pin was re-pinned to the background lane) — the resumed leg's ask and both rows' summaries carry `fromSubagent: true` (base: absent on the resumed ask and on both summaries).
|
|
300
|
+
- **Widening (open to veto @server @client-core).** The resumed leg's live `AskRequest.fromSubagent` goes from absent to `true` when a resume hands no parent task and the row records the bit (pinned: a host resuming a synchronous child's row with a bare config; the same join serves any leg whose row records the bit): the card says what the leg's first asks said. Who needs it: a supervisor that groups cards by origin; who is hurt: a consumer that read "no `fromSubagent`" as "a resumed child", none known; compensation: none needed — the record is engine-minted, never a TaskSpec seat.
|
|
301
|
+
- **Narrowing (open to veto @server).** A row that carries the bit owes the requirement word `from_subagent`: an engine before 7.68.0 refuses it `version_newer` before any claim (the row stays pending) instead of resuming the leg with its attribution dropped — the rolling-deployment cost every word member carries. A row whose `state.legTaskId` is present and not a string is `damaged` on every face (the member-shape table); a blank string reads (a host's `taskId: " "` is the id the leg's asks carried, so the row records it as minted).
|
|
302
|
+
- **Older rows.** A row minted before 7.68.0 records neither member: its summary carries no `fromSubagent` (nothing is inferred) and reads `agentId` as before (registry row, else session id).
|
|
303
|
+
- **Known limits.** KL-3212-a (#3212: a declared-`taskId` workflow agent's ask and summary named it two ways) is closed; `test/backlog3212-source-session-id.test.ts` P12's cell that pinned the two readings now pins one.
|
|
304
|
+
- **Rules.** Readings of the parked leg's task id: 2 → 1 on rows this release mints (the summary's lossy re-derivation stays only for rows without the record). Readings of `fromSubagent`: internals-only on a resumed leg → the one join every identity member a resume reads back uses. Added: two state members, one word.
|
|
305
|
+
- **Pins.** `test/backlog4083-fromsubagent-on-row.test.ts` — U1 the summary's literal-`true` echo · U2 `agentId` over the recorded task id (the older reading only without it) · U3 a damaged `state.legTaskId` on every face · U4 `from_subagent` rides exactly the rows that carry the bit · U5 the ask source's monotonic join · R0 a root run's row echoes neither · R1 · R2 · R3 (above) · R4 a blank declared task id parks a row this engine reads (a guard, green on the base). Nine red on the base (`d59ec91c6`), R4 green there; ten mutations each red a cell. Fences re-read: `test/backlog1320-reader-requirements.test.ts` (the word's maximal row, `legTaskId`'s one SHOWN read), the state-key fences in `test/design383-capture-optout.test.ts`, `test/durable-checkpoint.test.ts`, `test/steering-pending.test.ts`.
|
|
306
|
+
- **Consumers.** @server: read `CheckpointSummary.fromSubagent` and `agentId` as written — one projection, no derivation from `isDelegatedChild` (S-3920 / G51-b). @client-core: the summary's new optional member. @test: R1–R3 as black-box cells (a declared-`taskId` workflow agent's card vs its listing row; a synchronous child resumed by its host).
|
|
307
|
+
|
|
308
|
+
### A memory containment no longer puts the committed copy over bytes another writer landed while it ran — the inbound gate's put-back is the one every containment arm uses, and a put-back it declines is reported where theirs are (#4105; behavior change: a writer's newer edit of a file another read is containing is kept instead of replaced; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: +1 method `FileMemoryEngineBackend.drainInboundRestoreWithheld()`, additive; wire: none — the existing `HarvestReport.restoreWithheld` now also carries the inbound arm's rows, and a dual-root report now carries it from both planes) (@server @test)
|
|
309
|
+
|
|
310
|
+
- **Before.** When a read of a memory layer — the adopting read, a held layer's committed view (#4097), an audited restricted view — found an out-of-band edit that fails the inbound gate, it captured the bytes it read, took them off the plane, and then wrote the committed copy back by rename, unconditionally (`containInboundReject`, `src/core/memory-engine/file-backend.ts`). A writer that rewrote that file inside the window — after the read, or into the path the removal had just freed — lost its newer bytes to the committed copy, and no copy kept them. Since #4097 this is reachable in one process whenever another session of the layer is live: a live writer's next clean edit of a file whose gate-failing bytes another session's mount was containing was gone, and its harvest committed nothing (`liveWriterCommitted` 0 on `f9e774b75`). The `quarantine_failed` escalation for the after-read form said "the newer bytes … were left in place" while the put-back was overwriting them.
|
|
311
|
+
- **Now.** The put-back goes through `restoreContainedBaseline`, the restore the engine's containment arms already share: it overwrites only what the containment left at the path (its tombstone, or the suspect bytes when neither removal rung worked); a path somebody else holds keeps its bytes, and those bytes are judged on their own next read like any later write (a gate-failing rewrite is contained by that read, a clean one is adopted or stays the live writer's). The escalation's "left in place" sentence is now true. A put-back declined this way is recorded as a `HarvestReport.restoreWithheld` row (`{ path, entryId, rev }`, `rev` = the committed generation not written back) and reaches the host as the one `memory.restore_withheld` notice per pass, beside the engine arms' rows. A freed path is re-created at the store's projection mode (0o600, as before). Fixed on the way: a scope directory that refuses the removal tombstones the file in place and the committed copy now goes back over that tombstone (in place) — the rename left the tombstone standing; and `mergeHarvestReports` (dual root) dropped `restoreWithheld`, so the operator notice never fired for a dual-root deployment — it now concatenates both planes' rows.
|
|
312
|
+
- **Narrowing — open to veto (@server).** (1) An entry whose path is replaced by a symbolic link inside the containment window used to heal itself: the rename replaced the link with the committed copy. It now keeps the link, as every other containment arm does: nothing is served through the link (the link's target is never read), and the entry is not served at all until the link is removed or the entry is written again — its committed copy and its account row stay in the store, nothing is deleted. (2) Who is hurt: a reader of that one entry, between the window and the next ordinary write of it. (3) Compensation: the next harvest names the path (`symlink`), and the host's `memory.restore_withheld` names it with its entry id and committed generation.
|
|
313
|
+
- **Not changed.** The capture of the read bytes into quarantine, the removal rungs, the finding and its announcement, the retrieval tool's pure read, the engine arms' restores; callers still serve the committed copy for the pass that contained.
|
|
314
|
+
- **Rules.** Restore laws for a contained path: 2 → 1 (the backend's unconditional rename is retired; KL-3466-d / #3509 close with it). Drain sites of the backend's read-side record: 3 → 1 (`drainBackendInbound`, `src/core/memory-engine/engine.ts`). Faces that say "a containment did not put a committed copy back": the engine arms only → every arm.
|
|
315
|
+
- **Pins.** `test/backlog4105-contain-keeps-newer.test.ts`: three doors (adopting / held / restricted) × two moments (after the read, after the removal), the window injected at a fixed step — red on `f9e774b75` (the committed copy stands); per-door controls (nobody writes: the committed copy goes back); the newer bytes' own next read; the store's projection mode on a re-created path; the read-only directory's tombstone; the planted link; a live writer's harvest commits its newer note (engine); one `memory.restore_withheld` naming the path (runner) and its control; the dual-root merge. Re-pinned: `a005-memory-engine-nofollow` (the inbound restore is the shared restore — no-follow by both of its writers; the rename that replaced a link at this site is retired), `backlog783-containment-no-replace` ③ (the backend arm counted beside the engine's four; the free-path create takes the caller's mode).
|
|
316
|
+
- **Consumers.** @server: a host reading `HarvestReport.restoreWithheld` or `memory.restore_withheld` now also sees inbound containments (the out-of-session edit a read was containing); a dual-root deployment sees the notice at all. @test: the six window forms and the link form are deterministic injections in-process; a cross-process race is the black-box form.
|
|
317
|
+
|
|
318
|
+
### Errata to the 7.67.0 hard-breaking record, part 1 — fourteen type-level breaks the release's sections described without calling them hard-breaking, each said here in three sentences (#4138; found by the release post's item-by-item reading of the wire manifests against these notes; behavior change: none; narrowing: none; widening: none; hard-breaking: fourteen, three sentences each below, open to veto @server @cli @client-core @sdk @test; closed sets: none; exports: none; wire: none — no code changed, and the 7.67.0 block is what its tag carried; read against the types of `v7.66.1` and `v7.67.0` as the source declares them) (@server @cli @client-core @sdk @test)
|
|
319
|
+
|
|
320
|
+
- **Why this section.** The release's wire manifests (`wire-changes/<package>@7.67.0.json`) list 114 breaking items, and read by key the sections' hard-breaking sentences name 64 of them. The other 50 are answered in this and the next two parts: this part states fourteen breaks that no hard-breaking sentence stated (they answer 25 of the 50, and one more, `TaskSpec.workspace`, that a sentence names only by coincidence), part 2 names seven breaks the sections did state under other words (11 items), part 3 answers eleven readings that are not breaks (14 items). A hard-breaking sentence is one under a heading that carries `hard-breaking:`, so each of the fourteen below is stated that way now, for the release it landed in; `gate:changelog-breaking` holds every later release's notes to the same account.
|
|
321
|
+
- **HB-1 — `PeerHeldQueueSink.inject` may answer a promise (type level; #3878).** *What changed:* `PeerHeldQueueSink.inject` is `(entry) => void | "deferred" | Promise<void | "deferred">` where it was `(entry) => void | "deferred"` (`src/agents/peer-held-queue.ts`, a root type). *Who is affected:* code that calls a sink's `inject` itself — a wrapper or decorator around a sink — and reads the answer: a promise is neither `"deferred"` nor `void`, so a comparison with `"deferred"` still compiles and reads a deferral as a delivery; a host that only implements `inject` compiles unchanged, and the queue itself awaits the answer. *What to do:* `await` the answer of a sink's `inject` before reading it.
|
|
322
|
+
- **HB-2 — `TurnPromptSnapshot` has two more required members and a listener that takes the change (type level; #1393).** *What changed:* `TurnPromptSnapshot` (a root type) gains the required `declare(declaration)` and `mountedToolWireDigest`, and its `onChange` listener is `(snapshot, change: PrefixChangeNote) => void` where it was `(snapshot) => void`. *Who is affected:* a host or test double that builds a snapshot by hand (it lacks the two members), and code that calls `snapshot.onChange` itself (it must pass the second argument); a listener written for one argument still compiles and runs. *What to do:* take the snapshot from `buildTurnPromptSnapshot` rather than building one, or add the two members, and pass the change note where `onChange` is called.
|
|
323
|
+
- **HB-3 — `checkpointOccMatches` reads only `rev` from its expectation (type level, two narrow forms; #2515).** *What changed:* its second parameter is `Pick<ResolveExpectation, "rev">` where it was `ResolveExpectation` (`src/core/checkpoint-generation.ts`, a root function), and the function reads nothing but `rev` at both releases. *Who is affected:* a call that passes a fresh object literal carrying `holder` — `checkpointOccMatches(cp, { rev, holder })` now fails the excess-property check — and a replacement or mock typed as `typeof checkpointOccMatches` whose body reads `expect.holder`, which the parameter no longer has; a call that passes a `ResolveExpectation` value compiles as before. *What to do:* drop `holder` from the literal, pass the typed value, and read only `rev` in a replacement.
|
|
324
|
+
- **HB-4 — the store-contract kit reports a fourth optional member (type level; #3903).** *What changed:* the `onOptionalMember` callback of `toolResultStoreContract` is handed `{ member: "deleteBySession" | "listBySession" | "reap" | "putBytes"; status }` where `member` held the first three (`src/core/store-contracts/tool-result-store-contract.ts`); the kit now exercises the optional `putBytes` / `bytesOf` pair and reports its absence as `putBytes`. *Who is affected:* a store author's callback whose parameter is annotated with the old three-word `member` (it no longer compiles), and a `switch` over `member` that ends in a `never` default; a callback typed from the kit's own signature compiles and meets one more word. *What to do:* type the parameter from the kit's signature, or add the `putBytes` arm.
|
|
325
|
+
- **HB-5 — `workflowsCapability` reads only two `RunnerDeps` members (type level, two narrow forms; #3802).** *What changed:* its parameter is `Pick<RunnerDeps, "workflowScriptRunner" | "workflowGovernanceBaseline">` where it was `RunnerDeps` (`src/orchestration/workflow-script-runner.ts`, a root function), and the function reads those two members at both releases. *Who is affected:* a call that passes a fresh object literal carrying any other `RunnerDeps` member, which now fails the excess-property check, and a replacement or mock typed as `typeof workflowsCapability` whose body reads another member of `deps`; a call that passes a `RunnerDeps` value compiles as before. *What to do:* pass the value, or only the two members, and read only those two in a replacement.
|
|
326
|
+
- **HB-6 — three task row types carry a required `lastActivityMono` (type level; #3619).** *What changed:* `SemaTaskHandle`, `TaskListRow` and `BackgroundAgentTaskHandle` (which extends the first) gain the required `lastActivityMono: number`, the instant of the row's last activity on `performance.now()`'s clock, beside the wall-clock stamp the row already carried (`src/core/task-registry-shared.ts`, `src/core/task-registry.ts`). *Who is affected:* a test double, fixture or store that builds a handle or a list row as a literal; a handle or row the registry produces already carries it. *What to do:* register and list through the registry, or add `lastActivityMono` to the literal.
|
|
327
|
+
- **HB-7 — a handle's `updatedAt` is read-only (type level; #3619).** *What changed:* `SemaTaskHandle.updatedAt` and `BackgroundAgentTaskHandle.updatedAt` are `readonly`, because the registry moves it only together with `lastActivityMono`. *Who is affected:* code that assigns `handle.updatedAt = …` — a direct write is now a compile error; reads are unchanged. *What to do:* stop writing the member and let the registry stamp the row.
|
|
328
|
+
- **HB-8 — `ResumeRun` has a required `claimedRowPending` (type level; #3939).** *What changed:* `ResumeRun`, the second parameter of `Runner.runTaskStream`, gains the required `claimedRowPending: () => Promise<boolean>` beside `reopenUnstarted`: the row read that shows whether a reopen that threw had landed (`src/core/runner/contracts.ts`). *Who is affected:* a host that builds a `ResumeRun` itself and passes it to `runTaskStream`; `Runner.resume` / `resumeStream` bind it and are unaffected. *What to do:* go through `resume` / `resumeStream`, or supply a function that answers whether the claimed row reads `pending` at a generation past the one the claim won at.
|
|
329
|
+
- **HB-9 — a `prompt.snapshot_changed` trace frame has a required `cause` (type level; #1393).** *What changed:* the `prompt.snapshot_changed` arm of `TraceEvent` gains the required `cause: PrefixChangeCause` (and an optional `source`, for the `tool_table` word only) (`src/core/trace.ts`). *Who is affected:* a downstream that hand-builds a frame of this arm (a mock tracer input, a fixture); code that only reads a frame's members is untouched. *What to do:* give the hand-built frame its `cause`, or read frames from a real run.
|
|
330
|
+
- **HB-10 — a session head's `incarnation` is an optional member (type level, one narrow form; #3774).** *What changed:* `SessionHead.incarnation` and `SessionRowExpectation.row.incarnation` are `incarnation?: string | undefined` where they were required members that could hold `undefined` (`src/engine/harness/types.ts`); a head for a host-held row, or a row that predates the cell, now carries no such key. *Who is affected:* code that assigns a head, or an expectation's row, to a type that requires the key (`{ incarnation: string | undefined }`); reading `head.incarnation` still gives `string | undefined`, and a head literal that wrote `incarnation: undefined` still compiles. *What to do:* make the receiving type's `incarnation` optional, or build the value with the key.
|
|
331
|
+
- **HB-11 — two store errors carry one more code (type level; #1023).** *What changed:* `BackgroundAgentStoreError.code` is `"agent_record.already_exists" | "agent_record.row_malformed"` and `WorkflowRunStoreError.code` is `"workflow_run.already_exists" | "workflow_run.row_malformed"`, where each was its first word alone (`src/core/background-agent-store.ts`, `src/core/workflow-run-store.ts`). *Who is affected:* a `switch` over `error.code` that ends in a `never` default, a table keyed by the code, and code that types the code as the one old word. *What to do:* place the new word and treat it as the row's damage, not as a retry.
|
|
332
|
+
- **HB-12 — five places that carry a degrade reason read the one list `DegradeReason` (type level; #3567).** *What changed:* `degraded.reason` on `TaskResult` and on the three result types that extend it (`CascadeRunResult`, `RepairResult`, `VerificationResult`), and the `reason` of `TraceEvent`'s `task.degraded` row, are `DegradeReason`: the five words they carried (`breaker_open`, `rate_limit`, `budget`, `server_error`, `last_resort`) all remain, and `network`, `auth` and `stream_torn` arrive — so the wire manifest's reading of the `TraceEvent` rewrite as four lost members (`breaker_open`, `last_resort`, `rate_limit`, `server_error`) is an artifact: none is lost. *Who is affected:* a `switch` over `reason` that ends in a `never` default, and a table keyed by it. *What to do:* add the three arms.
|
|
333
|
+
- **HB-13 — `TaskSpec.workspace` may be `null` (type level; #3742).** *What changed:* `TaskSpec.workspace` is `WorkspaceDeclaration | null` where it was `WorkspaceDeclaration` (`src/core/task-spec.ts`); `null` is a leg that wants no tree. *Who is affected:* code that reads `spec.workspace` and hands it on as a `WorkspaceDeclaration` without narrowing `null`. *What to do:* handle `null` where the field is read.
|
|
334
|
+
- **HB-14 — a workflow baseline's `workflowModelAllowlist` may be the word `"catalog"` (type level; #3816).** *What changed:* `WorkflowGovernanceBaseline.workflowModelAllowlist` is `string[] | "catalog"` where it was `string[]`, and `"catalog"` follows the catalog the leg runs under (`src/core/runtime-caps.ts`). *Who is affected:* code that reads the field with an array-only method — `.map`, `.filter`, `.forEach`, `.join` — which no longer compiles without narrowing, and code that uses `.includes`, `.length` or an index, which still compiles but reads the word `catalog` as a string. *What to do:* read the names through the exported `workflowModelNamesOf(baseline, models)`, which answers for either form.
|
|
335
|
+
|
|
336
|
+
### Errata to the 7.67.0 hard-breaking record, part 2 — seven breaks the release did say, named here by the wire manifests' own keys so that a comparison by key finds them (#4138; behavior change: none; narrowing: none; widening: none; hard-breaking: seven items, each already stated in the section it points to, open to veto @server @cli @client-core @sdk @test; closed sets: none; exports: none; wire: none — no code changed) (@server @cli @client-core @sdk @test)
|
|
337
|
+
|
|
338
|
+
- **Hard-breaking — seven items, each stated above in the section named, under other words.** A manifest item names a member by the type that declares it; the sentences named it by the event, the seat or the family. Each line gives the key and the sentence.
|
|
339
|
+
- `ENGINE_NOTICE_CODES` loses `instructions.source_changed` (the `SB-2` line of the #3914 section: it was server-minted, and the lane's frame says what the notice said) and `mcp.revocation_probe_failed` (the `HB-2` line of the #3621 section: renamed `protocol.revocation_probe_failed`).
|
|
340
|
+
- `TaskEvent.source` loses `instructions_change` (the `HB-7` line of the #3914 section says it as `steering_injected.source`: `steering_injected` is the `TaskEvent` variant whose `source` that is).
|
|
341
|
+
- `FileSessionPolicyStore.putRules` and `InMemorySessionPolicyStore.putRules` take `SessionRulesInGrammar` where they took `SessionPermissionRules` (`HB-2` of the #1810 section: the interface and "both bundled stores").
|
|
342
|
+
- `WorkflowRunSummary` and `BackgroundAgentRowSummary` became unions (`HB (1)` of the #1023 section), and `TaskListStore.list` answers `TaskListEntry[]` (`HB (2)`): the damaged arm holds only the identity and `rowVerdict`, so a read of any other member needs `"rowVerdict" in summary` first. The members that moved into the readable arm are, for `WorkflowRunSummary`, `agentCount`, `agentFailures`, `budgetOvershoot` (`budgetTokens`, `spentTokens`, `unsettledTokens`), `createdAt`, `currentPhase`, `description`, `endedAt`, `errorCode`, `name`, `originatingSessionId`, `parked`, `phaseCount`, `startedAt`, `status`, `timeoutInterruption` (`agentsCompleted`, `agentsFailed`, `agentsInFlight`, `approvalWaitedMs`, `timeoutMs`) and `tokens` (`id` and `scope` stay on both arms); for `BackgroundAgentRowSummary`, `agentType`, `name`, `owner`, `parentSessionId`, `requestedBy`, `rootSessionId`, `sessionId`, `sessionScoped`, `settledAt`, `spawnedAt`, `status`, `updatedAt` and `usage` (`handle` stays on both arms).
|
|
343
|
+
- `MaterializedMcp` gains five required methods — `getPrompt`, `listPrompts`, `listResourceTemplates`, `listResources`, `readResource` (the #3870 section's heading: "a host that implements `MaterializedMcp` itself adds five methods").
|
|
344
|
+
- `PromptManifestSelectors` gains the required `locale` (`Hard-breaking ③` of the #2275 section, the manifest record).
|
|
345
|
+
- `TraceEvent`'s `brain.call` row has optional `cacheRead` and `cacheWrite` (the #1393 / #1119 section's hard-breaking bullet says it as "the `brain.call` trace row", beside the root `Usage`).
|
|
346
|
+
|
|
347
|
+
### Errata to the 7.67.0 hard-breaking record, part 3 — eleven manifest items that are not breaks, each answered with what the two releases' types say (#4138; behavior change: none; narrowing: none; widening: none; hard-breaking: none of these eleven — each is a reading of the manifest that does not hold, below; closed sets: none; exports: none; wire: none — no code changed; the manifest reader that makes these readings is #3971) (@server @cli @client-core @sdk @test)
|
|
348
|
+
|
|
349
|
+
- **Not breaking — eleven readings of the manifest, answered by the types of `v7.66.1` and `v7.67.0`.** The manifests list these as breaking because they compare a type's text, not what it admits; none removes a member or narrows what a value may be, and where a line says "still compile" it means a value written against the earlier type — one exception holds throughout: a value that already carried a member of a newly added name with another type no longer matches.
|
|
350
|
+
- `TaskRegistry.rollbackParkedClaim` answers `ParkedClaimRollback`, the same four words (`rolled_back`, `failed`, `lost`, `retry`) as the inline union it replaced: a name for the list (`PARKED_CLAIM_ROLLBACKS`), no member added or removed.
|
|
351
|
+
- `modelCostToPricing`'s `cost` gains an optional `cacheWriteLong`: every earlier argument still compiles.
|
|
352
|
+
- `maybeCompact`'s answer gains optional `restoreOutcomes` and `windowBelowOverhead`, and its `unevaluableWindow` member is now spelled `UnevaluableWindow` (the same members, `windowField` a `TriggerWindowField` = `autoCompactTokens` / `contextTokens` / `contextWindow`): code that reads the answer's members is unaffected.
|
|
353
|
+
- `ModSeatView` gains the required `turn`, but is not a root export and is reached only through the indexed types `LegFacadeSeats` takes from it (`exec`, `book`); the break a host can meet there is `LegFacadeSeats.turnAbort`, which the #3660 section states.
|
|
354
|
+
- `ReadInboxToolOptions.pull`: the function a host's `pull` hands back takes an optional fourth parameter `signal`, and the `PulledInbox` it answers gains an optional `notTaken`; a host function with fewer parameters, and an answer without the member, still compile.
|
|
355
|
+
- `isProcessLive` lives in `src/core/process-identity.ts` and takes `ProcessRecord` — the four members it took inline (`pid`, `startedAtMs`, `startedSinceBootMs`, `host`) plus an optional `pidNs`: every earlier argument compiles.
|
|
356
|
+
- `MemoryEngine.readDanglingSessionAccounts` answers `Promise<SessionAccountRow[]>` at both releases (it was written `Awaited<ReturnType<typeof foreignDanglingSessionAccounts>>`, the same type); `SessionAccountRow` gains an optional `holder`.
|
|
357
|
+
- `SyntheticWriteFacts` (not a root export) is `Omit<CallFacts, "budget">`: the five members it was — `face`, `cwd`, `root`, `advertisedWritableDirs` (with its `roots`) and `exemptRoots` — an interface rewritten as a type alias.
|
|
358
|
+
- `summarizeWorkflowRun` answers `ReadableWorkflowRunSummary`, the readable arm of the union and the members it answered before: narrower, so every reader of its answer compiles.
|
|
359
|
+
- `ConsolidationDirectedPatch.attribution` and `ConsolidationIntentCredentialRow.patches[].attribution` are `Exclude<MemoryAttribution, "overlap">`, which is `self` / `unattributed`, the two words they were: none is lost.
|
|
360
|
+
- `TaskListItem.status` is `(typeof TASK_LIST_STATUSES)[number]`, which is `pending` / `in_progress` / `completed`, the three words it was: none is lost.
|
|
361
|
+
|
|
362
|
+
### Errata to the 7.67.0 section for #1614 — the reference's hook input does name a source (#4138; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — no code changed; a sentence about the reference is corrected) (@cli @server @client-core)
|
|
363
|
+
|
|
364
|
+
- **What the section got wrong.** The #1614 section's bullet "The screen still runs" says the reference's submission-hook input "carries no source member". It does carry one: the reference's hook input schema declares an optional `source`, one of `user`, `sdk`, `system`, `loop_wakeup`, `schedule_wakeup` or `poll_event` (`cli321.js:276995`; the same enumeration at `cli316.js:261648`, so on 2.1.316 and 2.1.321), and describes `system` as the other machine-injected turns — peer messages, task notifications, auto-continuation.
|
|
365
|
+
- **What stays.** The reference still runs its submission hook on its own synthetic inputs (the coordinates that bullet gives are unchanged), and this release's screen still runs on every steer and names the source on its event: `e.source === "system"` is a word the reference's schema also has. Only the sentence about the reference changes; what the engine does does not.
|
|
366
|
+
|
|
367
|
+
### Errata to the 7.67.0 section for #3567 — the reference's fallback model holds for one user turn, not for the session (#4138; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — no code changed; a sentence about the reference is corrected) (@server @cli @client-core)
|
|
368
|
+
|
|
369
|
+
- **What the section got wrong.** The #3567 section's "Claude Code 2.1.316" bullet says the reference, once it has moved to a fallback model, "keeps the fallback for the rest of the session". The reference's own help for `--fallback-model` says it "Accepts a comma-separated list to try each in order. Re-tries the primary at the start of each user turn." (`cli316.js:280875`, `cli321.js:296489`, and the same option text at `cli316.js:67451`, `cli321.js:70382`): the chain holds within one user turn, and the next turn starts at the primary again. The reference's separate refusal fallback (`model_refusal_fallback`, listed at `cli321.js:276099`) is another mechanism and is not what that bullet described.
|
|
370
|
+
- **What it changes about the stated difference.** This engine's chain does not stick: every model request starts at the entry it names (D-3554-c), so a turn that makes several model requests tries the primary on each of them. Measured against the reference's hold for one turn, the distance is one turn's stickiness, not a session's, and the cost the section's sentence implied is smaller than it said. What the engine does is unchanged.
|
|
371
|
+
|
|
372
|
+
### Errata to the 7.67.0 section for #3618 — a background row waiting on a person at a live card is `running` and hears no beat, so a harvest window must not shrink to the single machine bound (#4258; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — no code changed; a consumer note is corrected) (@server)
|
|
373
|
+
|
|
374
|
+
- **What the section got wrong.** The #3618 section's consumer note says "@server: the harvest's "waiting" windows on a background row can fall back to the single machine bound, read with `lastActivityMono` (#3619)". It rests on the section's "Waiting on a person stays `status === "parked"`", which does not hold while a card is up on a live stream: the activity seat is told only at a call's `tool_start` and `tool_end` (`pushContent`, `src/core/runner/run-clock-and-content.ts:107-110`), at an MCP progress notification and at a shell's output frame whose totals advanced (#3638); a gate's question to a person is a turn inside the one tool call that is open, so the seat hears nothing while the person decides; and a background row turns `parked` only once its durable checkpoint exists, so while the card hangs on the live stream, before any durable park, the row reads `running` with no beat. Read in code; the row's status and beat count in that window are not yet measured on a real Runner (#4258).
|
|
375
|
+
- **What to do instead.** Until #4258 lands, a host's harvest window for session-held background rows stays no shorter than its own waits on a person (its live-card window, its approval wait) and its single-command ceiling — not the single machine bound. The cost of the shorter window, as the server reads it: a session-held background child calls a tool that asks, the person comes back after 30 minutes, a 10-minute window collects the row first, and the approval answers a run that is gone.
|
|
376
|
+
- **What stays.** The rest of that section holds: a foreground nested delegation's tool work and an MCP call that reports progress beat the row through the seat, and silence is read on the process's monotonic clock (`lastActivityMono`, #3619). What the engine does is unchanged.
|
|
377
|
+
|
|
378
|
+
### A turn's model-call wall time rides its usage — `turn_end.usage.apiDurationMs`, and the leg's total `TaskResult.stats.apiDurationMs`: a shell's "API time" is a sum of frames, no longer unknown (#4141; behavior change: one optional member on a frame's usage and one on the result's stats; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: `turn_end.usage.apiDurationMs?`, `TaskResult.stats.apiDurationMs?`, additive) (@server @cli @client-core)
|
|
379
|
+
|
|
380
|
+
- **Before.** The engine measured how long each model call took and wrote it only on the `brain.call` trace row (`latencyMs`); no frame and no result carried it, so a shell had no figure for the session's API time (the reference's result message carries it as `duration_api_ms`).
|
|
381
|
+
- **Now.** Each call's API wall time — from the instant the engine hands the request to the brain (the stamp the row already carries as `callStartedAt`) to the call's final message (the row's `ts`) — is summed where the call's tokens are: into the turn's usage (`turn_end.usage.apiDurationMs`, milliseconds, every call the turn made — a re-request inside a turn is added, not written over) and into the leg's stats (`TaskResult.stats.apiDurationMs`, the sum of its turns). The figure is the row's `ts − callStartedAt`: both instants are already on the row, nothing is added to it and no second clock is read, so the row, the turn and the leg agree to the millisecond. A call with no stamp adds 0: a hop a fallback walk left behind (a failover or model-lane walk is ONE call from the loop's side, so an abandoned hop's wall time sits inside the answer's row), a call beside the loop, and a turn no model answered (an extension's reply: a known zero, no `usageMissing`). The member is present on every `turn_end` and on every result of a leg that ran; a composition's stats (cascade, repair loop, verify) sum their legs', and carry none when no leg published one. Beside `usageMissing` it counts what was measured, like the token members. The leg's total is one run's, on the same footing as the token members: a session's total across runs is the host's sum, not an engine figure (@server @cli @client-core). The reference carries `duration_api_ms` on its final result message only (the including-retries total of its cost ledger, read from the corpus); the per-turn member is this engine's own, with no counterpart there.
|
|
382
|
+
- **Not changed.** Every member of the `brain.call` row, `latencyMs` (from the call's first delta; 0 for a call that streamed nothing), `firstTokenMs` and `callStartedAt` included, the usage members already on the frame and the stats, every budget and limit gate (the figure is observation only). The trace type's doc said two things of `latencyMs` ("from the brain request" and "anchors on the first delta"); it now says the second, which is what the code measures, and names `ts − callStartedAt` as the delivery-anchored one.
|
|
383
|
+
- **Known limits (disclosed, not defects; (1) is registered as D-4141-a).** (1) The reference's figure includes its retry waits (`cli321.js:287618`, `:275722`, `:293880`); this one counts each attempt the loop shows from that attempt's own issue to its end, the connection time, the wait for the first token and an abandoned failover hop included, and the waits between attempts not. (2) The side-call lanes (compaction, classifier, mods, advisor) are not timed (their rows write 0). (3) A delegated sub-run's time is its own result's `stats.apiDurationMs`; `stats.nested` carries tokens, turns, tasks and cost, not time. (4) A resumed leg reports its own calls: unlike `tokens`, no earlier slice's time is folded in (the resource ledger keeps no time). (5) A reply that carries no `usage` member at all (a host Brain's) writes no row and adds no time; a built-in adapter's usage-less reply carries a usage object flagged `usageMissing` and is timed like any call, and the turn's `usageMissing` is set when any call of it was flagged. A request the engine withholds and retries (a prompt-too-long recovery) is not seen by the account (its stamp is replaced by the retry's), so its time is in no row and no figure.
|
|
384
|
+
- **Rules.** Readings of a model call's wall time: 2 → 2, said apart (`latencyMs`, from the first delta; `ts − callStartedAt`, from delivery, the share of `apiDurationMs`), where the trace type's doc used to give both of `latencyMs`; no second clock. Special cases: 0.
|
|
385
|
+
- **Pins.** `test/backlog4141-turn-api-duration.test.ts`: T1 T2 T5 (a real Runner, a fake clock, two turns and one in-turn re-request, each call with a wait before its first content: the wire sum over the `turn_end` frames, the rows' Σ `ts − callStartedAt` and `stats.apiDurationMs` are one number, the wait before the first content is in it, `latencyMs` still reads from the first delta; the re-request is added to its turn; a failover walk's abandoned hop lands in the answer's row) red on the base (no member); T3 T4 the same for a cascade's and a repair loop's stats. `test/cumulative-stats-builder.test.ts` (the builder's every-field override and the tracker's sum carry the member); `test/backlog3571-prompt-submit-reply-s1.test.ts` / `test/backlog3572-prompt-submit-reply-s1b.test.ts` (the known-zero usage of a turn no model answered now spells the member, 0).
|
|
386
|
+
- **Consumers.** @server (the shell's `/cost` API-time cell reads the sum of `turn_end.usage.apiDurationMs`, or `stats.apiDurationMs` per leg; both are projections, nothing to derive). @cli @client-core: the same cell. On a `turn_end` absent means an engine that predates the member; on a result's `stats` it also means no leg ran here (a prepare-time refusal, a composition that dispatched no leg) — never 0.
|
|
387
|
+
|
|
388
|
+
### `isolationScopeOf` and `callScopeOf` are root exports — the two readings of a run's isolation scope, so a host that files rows under a run's scope asks the engine instead of copying the reading (#4027; server Q7151-S3799; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: +2 values `isolationScopeOf` / `callScopeOf`; wire: +2 root names, additive) (@server @sdk)
|
|
389
|
+
|
|
390
|
+
- **Before.** Both functions lived in `src/core/isolation-scope.ts` and the root named neither. A host that has to name the scope a run's row is filed under (a workflow run row's `scope` is what the Workflow tool reads through `callScopeOf(ctx, mounted)`, or the literal `default` when that answers nothing) could only deep-import the module or re-derive `ctx.isolationScope ?? mounted.scope` — a copy that keeps passing its own tests while the engine's reading moves, and that reads the empty string wrong the moment it is spelled with `||`.
|
|
391
|
+
- **Now.** The root exports `isolationScopeOf` and `callScopeOf`: the module's own bindings, re-exported, not wrapped. `isolationScopeOf(holder)` = `holder.isolationScope || holder.principal` (a blank member falls to the principal, a task without the member answers its principal verbatim); `callScopeOf(ctx, mounted)` = `ctx?.isolationScope ?? mounted.scope` (the context's resolved scope is carried as handed, the empty string included; the context's `principal` is never read). `isolationScopeRecordOf` (the member a parked row records) is deliberately not exported: nothing outside the engine reads it.
|
|
392
|
+
- **Not changed.** Both functions' bodies and every face that calls them. The snapshot's tier column reads `advanced` for both (its path heuristic, not a review).
|
|
393
|
+
- **Rules.** Added: none. Retired: none — two existing functions gain a root name, one reading each, no copy.
|
|
394
|
+
- **Pins.** `test/backlog4027-root-export-scope-readers.test.ts` — C1 both root names are the module's own functions (identity) · C2 `isolationScopeOf` through the root: member, principal, blank member, neither · C3 `callScopeOf` through the root: the empty string carried as handed, absent falls to the mounted scope, the principal never read · C4 a real TaskStop mounted at one scope stops a row exactly when the table says the context names that row's scope, and the root reader answers the table (five contexts × two row scopes, both arms present) · C6 a real Workflow launch records `callScopeOf(ctx, mounted)` as its run row's scope, the empty string carried, and `default` only when that answers nothing (six contexts) · C5 the snapshot rows, the consumer-smoke roster rows and the INTEGRATION-CORE sentence; `test/export-surface.snapshot.json` +2; `scripts/consumer-smoke-probe.mjs` registers both.
|
|
395
|
+
- **Consumers.** @server (Q7151-S3799): pin the workflow run row's `scope` against `callScopeOf(ctx, mounted)` from the root rather than the copy. @sdk: both names arrive in the root d.ts with the build.
|
|
396
|
+
|
|
397
|
+
### The judge of an approving person's written words is on the package root — `writtenFeedback`, `unfitFeedbackOf` and `UnfitFeedback` (#3973; server S-3725; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none new — `UnfitFeedback` is the three-word type #3854 minted, now nameable by a host; exports: +3 (`writtenFeedback`, `unfitFeedbackOf`, `UnfitFeedback`; tier advanced); wire: none) (@server)
|
|
398
|
+
|
|
399
|
+
- **Rule.** `src/core/approval-words.ts` is the one reading of "did the person write anything" (`writtenFeedback`) and the one judge of "may these words be delivered as the person's" (`unfitFeedbackOf`: on a refusal, by a machine, carrying the engine's `<system-reminder>` tag — in that order). The root re-exports the two functions and the type: the same bindings, not copies, with the same parameters and the same three refusals.
|
|
400
|
+
- **Before.** The two functions were reachable only inside the engine, so a host that takes the words at its own door (server's live `/respond`) either restated the judge or let the engine refuse after the host had already recorded a grant — an `allow_session` answer carrying a tagged text was refused by the engine for that call while the session grant was already written.
|
|
401
|
+
- **Now.** A host asks `unfitFeedbackOf(writtenFeedback(text), { allowed, byPerson })` before it records anything: a word back is the verdict the engine would refuse that answer on, on the run's approver seat, so the host refuses first and records no grant. The sentence a host writes for the refusal is its own (the engine's sentences are not exported); the verdict word is the engine's. Nothing about the engine's own reading changed.
|
|
402
|
+
- **Not changed.** The judge's semantics, parameters and the order of its three refusals; the live and durable faces' sentences and error codes; the answers a host that never calls the two functions gets.
|
|
403
|
+
- **Rules.** Added: none (three exported names over one existing reading and one existing judge). Retired: none. A host's own restatement of the judge is what this exists to retire — one judge, asked from both sides of the seam.
|
|
404
|
+
- **Pins.** `test/backlog3973-feedback-judge-export.test.ts` — G1 the root's `writtenFeedback` and `unfitFeedbackOf` are the same references as approval-words' (`===`), `UnfitFeedback` is the same three-word type · G2 each of the three refusal forms asked through the root is the answer the engine's own resolver refuses in (and a plain person's allow is fit on both) · G3 the frozen export surface holds the three names at the advanced tier; `test/export-surface.test.ts` 23 cells over the +3 with nothing removed.
|
|
405
|
+
- **Consumers.** @server (S-3725): call `unfitFeedbackOf(writtenFeedback(words), { allowed, byPerson })` at the respond door and refuse before recording the grant. The verdict agrees with a run that met the same answer (G2); the refusal's wording at your door is yours, the engine's own wording stays in its `approval.approver_defect` announcement.
|
|
406
|
+
|
|
407
|
+
### A mod handler is told who stands in the chain beneath it — `next.below`, the chain's members as the dispatch assembled it, so a consumer reads the list instead of matching an event against a matcher a second time (#4015; server S-3424; behavior change: one member on every `next` a handler and its `.catch` are handed; narrowing: none; widening: none; hard-breaking: one, type level, three sentences below, open to veto @server @client-core @sdk; closed sets: none; exports: +1 `ModParticipant` (type); wire: none — a handler face, not a wire frame, no key of any frame moves; the wire manifest reads the two required members as type-level BREAKING, which the sentences below say) (@server @client-core @sdk)
|
|
408
|
+
|
|
409
|
+
- **Rule.** What stands beneath a handler in one dispatch is answered by the one list the kernel builds when it assembles the chain (the participants of the event, matchers applied), read once as participants (`plugin`, `seat`, `observe`, the form a participant filter already sees); `next.below` is the part of that list after the handler, in chain order, and `next.trace` is the record of those members the run reached (a handler that answers without calling `next` leaves the ones under it unreached; they stay in `below`). A matcher is applied in one place; a handler that must know whether another handler stands beneath it reads the list.
|
|
410
|
+
- **Before.** A handler's `next` named only what had run beneath its last `next` (`trace`), after the fact; nothing said who stood beneath it before it called `next`, so a consumer that needed it (a hosted tighten-only lane refusing by site) judged the matchers itself, a second reading of every matcher.
|
|
411
|
+
- **Now.** `next.below: readonly ModParticipant[]` on the `next` of a handler and of its `.catch` (the same list — a `.catch` answers for its handler), frozen, in chain order, empty when nothing stands beneath. A handler the event's matcher leaves out is not in it; one the dispatch re-entered (asked through its `.catch`) is, as `trace` names it; a handler a `next.to` jump will pass over is, as `trace` names it `bypassed`. An observer is outside the chain — it runs before the chain, detached, and `trace` never names it — so it is beneath no one, and an observer's own `next.below` is empty. A task-seat member is carried with its seat; `observe` is `false` on every member of a handler's list (the field keeps the form a participant filter reads). The server's reading of a judging deployment handler beneath: `next.below.some((p) => p.seat === "deployment" && !p.observe)`. `ModParticipant` is exported (root and `@sema-agent/mods-kernel`).
|
|
412
|
+
- **Not changed.** The chain, its order, matchers, re-entry, the failure and budget law, `trace`, `next.to`; the participant filter a dispatch option takes; every wire frame and the closed sets (no key added).
|
|
413
|
+
- **Rules.** Added: none — one more reading of the participants list the dispatch already builds. Retired: none. The participant form is minted in one place (the filter and `below` share it).
|
|
414
|
+
- **Pins.** `test/backlog4015-mod-next-below.test.ts` — G1 a handler whose matcher the event misses is not in an outer handler's `below` (nor in its `trace`) and is when the event selects it; the server's reading is false when the only deployment handler is missed · G2 a task-seat member carries its seat; an observer, even one mounted after the outer handler, is not beneath it; every member of a list reads `observe: false` · G3 `below` and its members are frozen and a second read is the same list; the control reads each handler of a four-deep chain against its `trace` after the core · G4 a `.catch`'s list equals its handler's, empty for the last handler · G5 a handler re-entered through its `.catch` is a member of what is below it · G6 an observer's `below` is empty. Baseline red on 12 of 12; green here. Mutations M1–M10 each red on at least one cell, every cell red under at least one (not frozen · self included · matcher-blind roster · the roster from all participants, observers in · `.catch` given nothing · observer given the chain · a new list per read · re-entered members dropped · `observe` field wrong · the empty list unfrozen).
|
|
415
|
+
- **Hard-breaking — one, three sentences (type level, open to veto @server @client-core @sdk).** (HB-1) *What changed:* `ModNext` and `ModTypedNext` gain the required member `below` (it is not optional: the engine puts it on every `next` it hands a handler or a `.catch`). *Who is affected:* only a downstream that builds a `ModNext` / `ModTypedNext` value by hand — a test double handed to a handler, a fixture; a handler that receives `next` from `admitMods` or the engine reads it with no change, and no frame, key or closed set moves. *What to do:* give the hand-built value `below` (`Object.freeze([])` for a double with nothing beneath), or take the `next` from a real dispatch.
|
|
416
|
+
- **Consumers.** @server (S-3424): read `next.below` instead of matching a mounted handler's matcher; `ModMountedHandler.matcher` stays an admission record, not a second reading. @client-core @sdk: the d.ts gains the member as required; nothing in either repository reads or builds a `ModNext` (inferred: not scanned here).
|
|
417
|
+
|
|
418
|
+
### A parked approval card reads its why off the row the paused terminal names — the question, the askers' own words, the origin and the class pair are the row's `PendingAction.tool_approval` members, read by the terminal's `token`; the gate carries none of them (#3159; docs and pins only; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core @test)
|
|
419
|
+
|
|
420
|
+
- **Rule.** One home for why a call was put to a person: the ask's typed members on the live card (`AskRequest`) and their park twins on the row (`PendingAction.tool_approval`: `cardMessage`, `askReasons`, `origin`, `askClass` / `classes`), written by the park's one carry from the values the live card carried. The paused terminal (`terminal.kind: "paused"`) names the row — `token` is the store's key (`CheckpointStore.get`), `checkpointId` is the row's own — and its `gate` says which pause; the gate's `reason` is the engine's account of the pause kind, not the question. What the gate does carry about the ask is unchanged and is said beside the rule: its triage descriptor (`riskDescriptor`, with a probe's account when a probe tightened the call) and, on a real-approval ask, the bit the resume reads (`realApproval`, with the ask's origin word).
|
|
421
|
+
- **Why this is a reading, not a member.** The measured gap (a `rm -rf /Volumes/USB` park under `bypassPermissions` with `Bash(rm:*)` allowed showed no reason on the parked card) is a surface rendering the gate alone. The row carries the ask's `askReasons` entries (since 7.61.0; the removal guard's entry names it `integrity_guard` since #3934) beside `askClass` / `classes`; a copy on the gate would be a second store of one value. Written down on the terminal's `gate` member and in the integration guide's §4.6zf, where the stale sentence "`CheckpointSummary` does not echo [`askReasons`]" now says the summary echoes it (and still not the class pair).
|
|
422
|
+
- **Pins.** `test/backlog3159-park-frame-reads-row.test.ts` (a real Runner over a snapshotting remote workspace double; the drive is never touched — the call parks before execution) — F1 the attended leg: the terminal's token reads the row, the row's `checkpointId` is the terminal's, the row carries the guard's `integrity_guard` entry with the mounted-drive words, `askClass: "dangerous_delete"` first in `classes` and the card's question, each equal to the live card's for the same call, and the inbox lists the row under the same `checkpointId` · F2 the unattended leg (no approver, no durable opt-in): the same row reading · F3 none of the five members (nor a `message` / `askMessage`) is on the gate, and the gate is the row's gate.
|
|
423
|
+
- **Consumers.** @server: the parked card (the `suspended` frame's card, the approval inbox's parked rows) reads the row by the terminal's token — `cardMessage`, `askReasons`, `origin`, `askClass` / `classes` — instead of rendering the gate alone (the server half of this ticket; the same fact as its own deletion-guard sighting). @cli: the parked card renders `classes` and the entries' words from what the host projects off the row. @client-core: the parked-card projection takes the same members off the row the live card takes them off `AskRequest`. @test: the parked deletion-guard card carries the guard's sentence (the s139c parked-card cell) once the host reads the row.
|
|
424
|
+
|
|
425
|
+
### A userinfo with nothing in front of it whose password opens with a quote is read when the quote's partner stands right before the `@` — `admin:"hunter2"@db:5432`, `admin:'hunter2'@db/app` (#3806; behavior change: narrowing — the form is withheld on egress, by the scan's wide grade, by the route adjudicator's URL strip and from a child's environment, one sentence below, open to veto @server @cli; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli)
|
|
426
|
+
|
|
427
|
+
- **Rule.** The `url-userinfo` row's frontless form refused every password opening with a quote, so a quoted password and a quoted value around `admin@db:5432` were one shape. Now the quote is the password's own when its partner stands right before the `@` that ends the userinfo (no `"`, white space or `/` between the two quotes; that `@` read with its host, so the userinfo never ends at an `@` inside the quotes); any other quote still opens a quoted value — its partner after that `@`, anywhere else, or nowhere. One clause, the same row, so every face that reads the vocabulary answers alike; no reading is added.
|
|
428
|
+
- **Narrowing (open to veto @server @cli).** Before: `admin:"hunter2"@db:5432`, `admin:'hunter2'@db/app`, `admin:"hunter2"@[::1]:5432`, `connect admin:'hunter2'@db.example:5432 now` and `admin:"hun@ter2"@db:5432` passed egress, the wide scan and the URL strip as written, and the env predicate handed them to a child (the predicate of 7.66.0, before the one userinfo reading, withheld the four that open the value). Now: each is withheld on every face — the password marked on egress (`admin:«redacted»@db:5432`; a userinfo holding a second `@` whole, `«redacted:userinfo»@db:5432` — the marks of #3970 below), the userinfo marked by the strip, the variable withheld as `value-rule`. Who notices: a host scanning or rendering text with a `name:"word"@host:port` or `name:'word'@host/path` in it — a log line naming a user (`login user:"root"@10.0.0.5:22`), a code fragment (`k:"v"@h:1`), an attribute or a JSON string quoting one (`title="author:'jane'@example.com/profile"`) reads the same as a quoted password and is now marked too.
|
|
429
|
+
- **Not changed.** `{host:"admin@db:5432"}`, `{host:'admin@db:5432'}`, `connect(host:"admin@db/app")`, two quoted values in one object (`{a:"x@db:1",b:"y@h:2"}`), a quoted value followed by a bare `@` (`admin:"x@db:1",y@h:2`, ``{host:"admin@db:5432"+`y@h:2`}``) stay clean on every face; a password behind a scheme, an `https:` or a `//` is read as before; the env predicate's other KL-3769-a forms are unchanged. Left unread, in KL-3769-a: a quote with no partner (`u:"pw@db:5432`, locally the same as a cut-off quoted value), a partner followed by more of the password (`u:"pw"x@db:5432`, `u:"a"b"c"@db:5432`, `u:"pw",x@db:5432`) and a `"` inside a single-quoted one (`u:'p"w'@db:5432`).
|
|
430
|
+
- **Pins.** `test/backlog3806-quoted-userinfo.test.ts` — G1 the five forms on the five faces and the env entry on the public path (red before) · G2 the three quoted-value controls clean · G3 the fronted forms read as before · G4 the KL-3769-a row (red before) · G5 twenty-eight counter-examples each pinned to its env, wide and strip answer (seven red before: five accepted over-reads and two credentials now read) · G6 six long forms of the clause at 5.6 million code units within a CPU bound on four faces. Mutations each red, 6 of 6: the old clause back (G1, G5), any quote read as the password's (G2, G5, and #3769's K7 controls), the host check at the partner's `@` dropped (G5), more of the password allowed after the partner (G5), a `"` allowed inside a single-quoted run (G5), white space allowed inside the quotes (G5).
|
|
431
|
+
|
|
432
|
+
### The WebFetch summarizer receives its model at the model door — a Model carrying `fallbacks` that is not the catalog's own entry is refused when the summarizer is built, where its request walked no chain and said nothing (#4062; behavior change: `createWebFetchSummarizer`'s `model` is received as the Model a task, a role or a tier names is; narrowing: two, below and open to veto @server @test; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @test)
|
|
433
|
+
|
|
434
|
+
- **Finding.** The chain of a model request is read off the Model the catalog door received for the entry (`modelFallbackChainOf`, keyed by that snapshot's identity). `createWebFetchSummarizer(seat, model, …)` took `model` as the caller's object, with no door: a copy of the catalog's entry, or the object the host had written into its catalog, carried its `fallbacks` words and was asked on the Brain with no chain behind it — a rate-limited first model was the summary's failure, the declaration walked by nothing, no refusal anywhere. The 7.67.0 note's instruction to @server ("the catalog's `fallbacks` then reach the summary") held only for the catalog's own entry.
|
|
435
|
+
- **Now.** The summarizer receives `model` at the model door (`receiveModel`), when it is built, after the seat and before the model's own budget. The catalog's own entry (the Model a Runner's catalog holds under its word, or what `resolveTaskModel` answers over that catalog) comes back as itself and walks its chain. A Model with no `fallbacks` has no chain, as before. A Model that carries `fallbacks` and is no such entry is refused, `config.model_fallback_invalid`, naming `createWebFetchSummarizer's model` and the entry to pass. The request is asked on the door's snapshot; `modelOriginOf` answers the caller's object.
|
|
436
|
+
- **Not changed.** The chain walk, the seat, the policy gate, the per-call auth seat and the budget law; a catalog-entry snapshot is handed to the Brain as the same object it always was. The resolver over a catalog the caller built answers the door's snapshot of the entry with its chain recorded (#4150, shipped in 7.67.1; before that it refused an entry that declares `fallbacks`), so what it answers is such an entry and walks (pinned as a control).
|
|
437
|
+
- **Who is hurt (narrowing, open to veto @server @test).** (1) A Model carrying `fallbacks` that is a copy of the catalog's entry, or the object the host wrote into its catalog: refused at construction where it was built and walked nothing. (2) A Model the door cannot read — a member outside the `Model` contract, a retired key, a class instance, a Proxy, an `id` or `api` that is not a string, a value that is no object — refused at construction (`config.unknown_role`, `detail.reason: "not_a_model_ref"`, the seat named) where it was accepted and handed to the Brain as written: the refusal every other Model seat of the engine already makes. (1) never walked a chain, so no request that worked changes its answer. A host that keys its own table by the Model object it handed in reads `modelOriginOf(model)` on what the Brain and `getApiKeyAndHeaders` receive, as at every other seat. **Compensation.** None needed beyond the refusal's sentence, which names the entry to pass.
|
|
438
|
+
- **Rules.** Readings of "which Model a summarizer asks": 2 → 1 (the caller's object, and the door's snapshot, become the door's snapshot). Special cases: 0.
|
|
439
|
+
- **Pins.** `test/backlog3682-webfetch-summarizer-lane.test.ts` E1 E2 E5 E5b E6 E7 red on the base (no refusal; the Brain and the auth seat are asked on the caller's object; the model is not judged before the budget); E3 (a real Runner's catalog entry walks the chain) is a control, green on both; E4 (the entry the resolver answers walks, over a received catalog and over a caller-built one) is green here, its caller-built half refused on 7.67.0 (#4150 changed the resolver).
|
|
440
|
+
- **Consumers.** @server (S-3300 / PICK768 — the summarizer assembly of the 7.148.0 line): pass the catalog's own entry as `model` (what `resolveTaskModel(…)` answers over the catalog the Runner received, or the Runner's catalog entry), not a Model object the lane built or copied; a summarizer model that declares `fallbacks` and is not that entry is refused when the summarizer is built, the cause named. Whether the server's lane ever passes such an object is not established here: its main branch calls `createWebFetchSummarizer` with the model `resolveTaskModel` answered over its own configuration table (read, not run). @test: one cell for the refusal — a copy of a catalog entry carrying `fallbacks` handed as the summarizer's model answers `config.model_fallback_invalid` at construction.
|
|
441
|
+
|
|
442
|
+
### A copy of another version of a file ends the chains of the file in the request, the seed's among them — a file changed outside every write call and changed back is sent whole where the request said "unchanged since you last read it" (#4061; behavior change: the third Read of a file read as X, then as Y, then as X again carries X's contents in the request, where it was the sentence that points at the first copy; narrowing: none; widening: one, below and open to veto @cli @test; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @test)
|
|
443
|
+
|
|
444
|
+
- **Finding.** The request's copy projection ended a chain of copies only at a write call (`Edit` / `Write` / `NotebookEdit`) to the file's name. A file changed by a shell command, an editor or a checkout — X, then Y, then X again, no write call between — left the first copy's chain standing: the third Read, X again, joined it, and the request rendered `[f.md: unchanged since you last read it (lines 1-3 of 3); content omitted to save context]` after the model's last sight of the file had been Y. The leg's seed already ended on another version of its file, by a second rule beside the write call's.
|
|
445
|
+
- **Now.** One law. A copy that repeats no standing chain is another version of every chain of its own view that holds one of its spellings, and ends them (as a write call to the file's name does); it heads a chain of its own, so the third copy carries its contents (a frame too) and the next copy of it repeats it. A VIEW is what of the file a copy shows: the whole text, one window (a range), a notebook as a Read, a notebook as a frame — copies of one view with other contents are two versions; copies of different views say nothing of each other's version and end nothing. The leg's seed is the head of its file's whole-text chain, in place before the view's first message, holding its canonical key alone (a spelling reaches it by file name, a chain by the exact spelling its copies were delivered under, as before); it ends the way every chain ends. A copy that repeats a chain, a seed included, ends nothing.
|
|
446
|
+
- **Not changed.** The first copy's contents, the "not repeated" sentence beside its copy, the pointer sentence, the budget arm, the withdrawal frame's end, the retired-sentence readings, the changed-files notices, a transcript. Another spelling of the path is another file's copy (pinned): a Y under `/w/b.md` does not end `/w/a.md`'s chain. A whole copy, a copy of line 2, a whole copy again — the third still repeats the first; window A, window B, window A likewise; two seeded files of one file name copied A, B, A — all three are their seed's. What a repeat renders is #4180's (this release, below): a frame that repeats is the "not repeated" sentence, or the seeded one; a `Read` result is its contents on every request (`CopyWalk.decide`).
|
|
447
|
+
- **Who is hurt (widening, open to veto @cli @test).** One: a window read of a seeded file leaves the seed standing. The seed used to end on any copy of another identity under its file name, a window of the file included; a window cannot be told from a version of the file, so it now ends nothing, and a later whole copy of the seeded contents answers the seeded sentence — true, because a whole copy equal to the seed's hash is the seed's bytes on disk. **Compensation.** None needed. Still open and not changed here (the same on the base): a token-capped delivery is no copy and ends no chain; a frame that joins one of two chains of one identity under two spellings leaves the other chain standing.
|
|
448
|
+
- **Rules.** What ends a copy's standing: 3 → 2 (a write call ended the chains by one mechanism and the seed by a second name set; another version ended the seed by a third per-key set; now one `end` over the chains, the seed among them — the seed's two sets are retired). Special cases: 0 (the seed's file-name reach is the one reading of "holds the spelling"; the view is part of a copy's identity).
|
|
449
|
+
- **Pins.** `test/backlog3904-read-state-s1b.test.ts`: red on the base — V1 (Read X, Read Y, Read X), V2 (a frame of X after them), V7 (a real run's third Read), V12 (a frame of Y between two Reads of X), V8 and V9 (their last assertion: the same window, a notebook, with other contents between); green on both — V3 (another spelling's Y), V4 (X, X, Y), V5 (the seed ended by another version, a Read spelled by its file name), V6 (a write ends the seed), V10 (two seeds of one file name), V11 (a frame's alias label is no spelling of the seed), and V8 / V9's other assertions (a window or a notebook's other carrier between two copies of unchanged contents ends nothing). Fifteen mutations of the law's arms and the door's, each killed by a named cell.
|
|
450
|
+
- **Consumers.** @cli @test: nothing to change. @test: one cell — a file read, rewritten outside the agent's tools, read, rewritten back, read: the last request's third result is the contents, not the sentence.
|
|
451
|
+
|
|
452
|
+
### A parent parked on an approval and the background agents it launched — the yield-and-continue form of #2273 answers #1097, now pinned per case a person sees (#1097; no behavior change — covered by #2273, guard pins and two known limits; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @server @test)
|
|
453
|
+
|
|
454
|
+
- **Rule.** A task-owned background agent belongs to its owner task, not to the leg: a leg that durably parks (any gate) yields the agents it owns that can park and stops the ones that cannot, before its result is published; the owner's resumed leg continues what yielded; a leg that ends its task stops them. A session-scoped agent belongs to its session and is not the parked leg's to decide.
|
|
455
|
+
- **Read, not changed.** #1097 reported that a parent parked on an approval had its background agents reaped by the teardown. That held until #2273 landed; since then, with a durable row store, the checkpoint store and the session-durability capability, the agent parks beside its parent and continues on the parent's resumed leg — in the same process or another one over the same stores — and the parent's resumed conversation hears it finish once. An agent that finished inside the batch its parent parked on, and a session-scoped agent that finished while the parent was parked, are each heard once when the parent resumes in the same process.
|
|
456
|
+
- **Known limits.** KL-1097-a: while the parent waits on the person its task-owned agents do not progress (the reference keeps background agents running through a parent's permission wait); an agent that cannot yield is stopped and named once on `onError`; across replicas, a session-scoped agent keeps running in the process that spawned it and a resume elsewhere neither sees it live nor receives its completion unless the deployment parks engine frames in a durable session box (`RunnerDeps.sessionBoxEngineFrames`), and a task-owned agent that finished inside the parking batch is not heard by a resume in another process at all — its frame, still on the parked leg's lane, is held in the parking process (#4057 routes that end-of-run park through the one park seat). KL-2273-a: an agent whose final answer is in flight when the yield lands parks at the boundary after that answer and answers again when continued.
|
|
457
|
+
- **Pins.** `test/backlog1097-bg-subagents-park.test.ts`: M1 / M2 (same-process and cross-process resume — red on the tree before #2273, green now), M3a / M3b (a child finishing in the parking batch, a session-scoped child finishing during the park; same-process resume), K1 (KL-2273-a), K2 (KL-1097-a, the parking-batch child with a cross-process resume) and K3 (a yielded child's poll still says it waits on an approval decision under a state line that says yielded; #4066) recorded.
|
|
458
|
+
|
|
459
|
+
### The person is told when a run ends without reading another author's accepted input — the engine notice `task.undelivered_input_dropped` (audience `user`, once per session, `detail.authors` counting per author), beside the operator's `onError` line, which stays (#3954; client-core CC-1069 / AU-4; behavior change: one notice row a run that drops such an input now sends, where only the operator's sink heard before; narrowing: none; widening: one, below, open to veto; hard-breaking: none; closed sets: `EngineNoticeCode` +1, `task.undelivered_input_dropped` (audience `user`, minter `engine`); exports: none at the root; wire: one notice code, `detail: { authors, sessionId, sourceSessionId?, agentId?, taskId, runId }`; the notice's audience, open to veto) (@server @client-core @cli @test)
|
|
460
|
+
|
|
461
|
+
- **Before.** An observer's report is an input another author wrote into the observed child's run. When the child's leg ended before any turn read it, the report was dropped (never parked, never on the person's undrained account) and the loss reached only the host's `onError` (`phase: "degraded"`, classification `undelivered-input-dropped`) — a server turned that into an operator log line, and a client's wire carried nothing, so the person watching never learned that part of the observation never arrived.
|
|
462
|
+
- **Now.** The same sink that feeds `onError` also mints `task.undelivered_input_dropped` (`droppedInputsNotice`, `src/core/input-author.ts`; wired in `src/core/runner/run-identity-wiring.ts`): the sentence names how many inputs each author wrote that the leg ended without reading, `detail.authors` counts per author (only authors that wrote one are keys). It is told ONCE PER SESSION — the digest rides the leg's own announced record, keyed on that leg's session id, and the write is parked for the leg's exit flush; the cap is best-effort in the one direction the announced record always errs (a write the exit could not persist, an older park mirror a resume reads, a second Runner on the session, or a record past its bound can tell a later loss again; none drops a first one). It is addressed like every user-face notice of a leg (`Prepared.legNotice`, `noticeAddressOf`): a top-level leg's own session; a delegated leg's tree root, with `detail.sourceSessionId` / `agentId` naming the leg — one row, never a second one on the spawner's session. The `onError` line is unchanged (once per leg end): one loss, two faces, each told to its own audience.
|
|
463
|
+
- **Widening (open to veto @server @client-core).** The person's wire gains one `user`-audience notice a run that drops another author's input sends; a host that routes `user` notices by `ENGINE_NOTICE_CODES` / `NOTICE_AUDIENCE` receives it with no change, a host with its own whitelist adds the word.
|
|
464
|
+
- **The audience (open to veto; a choice of this engine's, not a departure from the reference).** Audience `user`, not `operator`: the loss is what the person's session was told and never read, and what to do about it (nothing to re-send — the observer's next segment reports anew) is theirs; the operator's face is the line that stays.
|
|
465
|
+
- **Not changed.** The `onError` line, its classification and its once-per-leg-end count; the person's own undrained account (`task.user_steer_undrained` / `task.user_followup_undrained`) — an operator's input never reaches this notice, and another author's never reaches that account; what a pause carries (only the operator's input).
|
|
466
|
+
- **Closed sets.** `ENGINE_NOTICE_CODES` +1 (the release's count, 135 → 153, is in the #3832 section): `task.undelivered_input_dropped`, audience `user`, minter `engine` (`docs/CLOSED-SETS.md`, `docs/INTEGRATION-CORE.md`, the row on `EngineNotice.code`); `docs/KNOWN-LIMITS.md` KL-2941-b now names the notice.
|
|
467
|
+
- **Rules.** +1 catalog word with one row in each of the two tables; no new door, no special case: the sink the harness already calls is the one mint site, and the address is the leg's existing notice route. The sibling undrained family's copy-to-parent is the older addressing form and is untouched.
|
|
468
|
+
- **Pins.** `test/backlog3954-undelivered-input-notice.test.ts` — G1 a leg ending with an observer's report ⇒ one notice, closed-set word, per-author counts (two authors, two inputs of one), a delegated leg ⇒ one row addressed to the root naming the leg · G2 the same session's second drop is not told again while the operator's line still is, a fresh session is told, the digest is keyed on the leg's own session · G3 a report the run read ⇒ nothing, the person's own undelivered steer ⇒ its own account and never this one · G4 the word is catalogued once, audience `user`, minter `engine`, written in CLOSED-SETS and on the `EngineNotice.code` rows; the catalog-listing pins (`governance-codes`, `backlog1279-child-undrained-to-parent`) name it.
|
|
469
|
+
- **Consumers.** @server: add `task.undelivered_input_dropped` to the user-notice whitelist (the half-field your side owns). @client-core: message catalog, two languages; the tripwire `run-input-author-projection-test` T segment fires on this word. @cli: render it where the other `user` notices are rendered. @test: black-box — an observed child whose report is dropped at leg end shows one notice on the tree root's stream.
|
|
470
|
+
|
|
471
|
+
### A parked agent's card is read through the store's body-tolerant listing when its body does not decode, and a card nobody can read is announced once — a voided card with a lost body no longer leaves its agent row `parked` for good (#3983; behavior change: a parked row whose card cannot be decoded but whose stored status is `expired` is failed by the next sweep; hard-breaking: none; closed sets: one new, `ParkedCardState` (`pending` / `resolved` / `expired` / `missing` / `unreadable`, the answer of the one card-state reader); exports: none) (@server S-3740)
|
|
472
|
+
|
|
473
|
+
- **Rule.** One function, `parkedCardState`, answers what state a parked row's approval card is in. The registry's live sweep and the store sweep (both of its arms) ask it, and nothing else reads a card's state. The body read answers when the body decodes; when it throws, or the card carries no state word, the store's session listing answers (`CheckpointStore.listBySession`, whose unreadable rows carry the status as stored); a card neither read can speak for is `unreadable`, which the sweeps treat as they always have (touch nothing) and which is announced once per scope and agent per process.
|
|
474
|
+
- **Before.** Each half read `get(token).status` itself — two readers of one fact — and caught any throw as "store outage: touch nothing this sweep". A store that splits a card into a state column and a body (the SQL twin throws from `get` for a body that is not a checkpoint, a JSON `null` among them) therefore made a voided card with a lost body read as an outage for ever: its agent row stayed `parked`, silently, with nothing for any sweep to act on and nothing said.
|
|
475
|
+
- **Now.** A readable card costs the one `get` read it always did and never touches the listing. A card whose body does not decode is read through the listing: stored status `expired` fails the row (`approval checkpoint expired while parked (reconciled)`), `pending` leaves it (the card's own deadline ends it; the sweep does not void what it cannot prove voided), `resolved` goes through the claim and re-read guards it always had. The listing is read page by page to its end, and given up on only when a page has no rows or its cursor did not move past the one asked with. A card no read can speak for leaves the row parked and says so once, naming the agent and never the card's token (a resume capability).
|
|
476
|
+
- **Not changed.** What each arm does with each state; a readable card's read; the sweeps' names and arguments (the store seat widens by one optional member, `listBySession`, which a store that lacks it simply does not offer); the resolved arm's re-read, which now asks the same reader.
|
|
477
|
+
- **Not breaking — the sweeps' store parameter (a wire-manifest reading).** `reconcileParkedAgents(stores, …)` and `TaskRegistry.reconcileParkedAgents(stores, …)` take `stores.checkpointStore` as `ParkedCardStore` = `Pick<CheckpointStore, "get">` and an optional `listBySession`, where it was `Pick<CheckpointStore, "get">`: every earlier argument compiles (a value that carries a `listBySession` of another type is the one that no longer matches).
|
|
478
|
+
- **Not fixed here (disclosed).** A store that does not offer `listBySession` — the SQL twin does not at server 7.148.0 — cannot state a lost-body card's status, so its row stays parked and the sweep now says so once. For that store the row converges when it lists unreadable rows by session, or when the contract gains a status read that does not decode the body (a store-contract change, not made here).
|
|
479
|
+
- **Pins.** `test/backlog3983-reconcile-unreadable-card-body.test.ts` — G1a a really parked agent whose card is voided and body lost is failed by the live sweep, G1b by the store sweep over another process's row · G2a a readable expired / pending card is read with one `get` and no listing, G2b a missing card still fails the row as missing · G3a a card no read can speak for leaves the row parked and warns once across two sweeps without the token, G3b a listed status outside the three, G3c a pending card with a lost body is not voided and not announced · G4a a voided card beyond the first listing page is found, G4b a store whose cursor never moves past the page it was asked for is read twice and then given up on. Mutations (receipt).
|
|
480
|
+
- **Rules.** Retired: the two inline card-state readers (the live sweep's, the store sweep's closure) and the resolved arm's separate re-read. Added: one reader with a fixed order (body, then listing) and one announce-once latch. Net: one reader instead of two.
|
|
481
|
+
- **Consumers.** @server (S-3740): nothing to change for a store that already offers `listBySession` (the local checkpoint store does, through core's); for the SQL checkpoint store, offering it (rows it cannot read listed `unreadable` with the stored status, never skipped, as the contract states) is what lets a sweep settle a row whose card's body is `null`. No wire key or export changes; the one new closed set is `ParkedCardState` (above). The "announced once" line rides the operator notice `delegation.parked_card_unreadable`, which #3832 in this release adds (below) — a notice router keyed by `EngineNoticeCode` gives it a row there.
|
|
482
|
+
|
|
483
|
+
### A parked ask's gate reason says what the ask is — a question waits on an answer, any other call on an approval, and only a tool's own safety mark says "irreversible" (#3991; behavior change: one text member, `gate.reason`, on the two plain gate kinds — `human` and the unattended `irreversible_ask`; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — `gate.reason` is an existing free-text member, its text for the asks below changes) (@client-core @cli @test @server)
|
|
484
|
+
|
|
485
|
+
- **Rule.** A parked row's `gate.reason` says what the ask IS. The two plain gate kinds — `human` (the opt-in kind) and `irreversible_ask` with no safety mark (the unattended kind) — word an ask as what it is: a content ask (the engine's question tool, judged by `isEngineQuestionFace`, the park's own judge) waits on an answer, `AskUserQuestion: awaiting a human answer (durable)` (the sentence the durable-question policy asks with); any other call waits on an approval, `human approval required for tool "<tool>"`. A question parks only where the leg has a durable opt-in, so the opt-in kind is the one that files it and the unattended kind words an approval only. "Irreversible" is said only by the safety arm (`human approval required before safety-tightened tool "<tool>"`), which reads the tool's own mark; a marked tool never reached the unattended arm.
|
|
486
|
+
- **Before.** The unattended kind's one literal covered every ask it served — `human approval required before irreversible tool "<tool>"` — every unmarked tool the policy asked about included (a root leg's plain ask parks there when a mounted irreversible tool armed the lane; a question does not reach this kind since 7.67.0 — it parks only on a leg with a durable opt-in, so it is filed under the opt-in kind). The opt-in kind called a question an approval, `human approval required for tool "AskUserQuestion"`. Beside the card the row already said `approvalCard: "question"` for the question.
|
|
487
|
+
- **Old → new, the two pairs.** A question, on the opt-in kind: `human approval required for tool "AskUserQuestion"` → `AskUserQuestion: awaiting a human answer (durable)` (the unattended kind files no question, see Before). Any other call on the unattended kind: `human approval required before irreversible tool "<tool>"` → `human approval required for tool "<tool>"` (the opt-in kind's approval sentence, unchanged).
|
|
488
|
+
- **Now.** `durableQuestionAwaitingSentence(toolName)` (`src/core/ask-question.ts`; exported from its module, not from the package root) is the one spelling of the question sentence: the durable-question policy asks with it and the opt-in kind's park files it as `reason` when `contentAsk`, an approval sentence otherwise; the unattended kind's `reason` is the approval sentence. `durableApprovalAwaitingSentence(toolName)` beside it is the one spelling of that approval sentence — both arms call it. The gate kinds, the card (`approvalCard`), the risk descriptor, the scope, the deadline and the pending call are what they were; only the text of one member moves. The card's own `cardMessage` is the asking policy's words as before (a card whose policy spoke differently, or that carries an allow-rule explanation, keeps those; the row's reason does not follow them).
|
|
489
|
+
- **Not changed.** The kind of every arm; the safety arm's and the real-approval arm's words; the opt-in kind's words for an approval ask; rows parked before this change keep the reason they were filed with (nothing rewrites a stored row; a search of `src` finds no reader of a parked gate's reason).
|
|
490
|
+
- **Rules.** Retired: the unattended arm's reading of every ask as an irreversible tool; the opt-in arm's reading of a question as an approval. Added: one function, `durableApprovalAwaitingSentence`, the approval sentence both plain arms already wrote (a second spelling retired).
|
|
491
|
+
- **Pins.** `test/backlog3991-park-ask-reason.test.ts` — G1 a root leg whose facility is the entitlement alone files the question under the question sentence on kind `human` (and, with the durable-question policy asking, the card's message) · G1b the same from a background child on a park-capable lane, once holding no marked tool, once holding an egress tool · G1c where no question parks (a root leg with a store and an armed vocabulary but no opt-in; a synchronous child, both shapes) no row is written and the sentence the model reads names no approval and no irreversible tool · G2 an unmarked tool the policy asks about reads the approval sentence on the unattended kind, no "irreversible" · G3 a declared opt-in files the same question under the question sentence, kind `human` · controls: G2b a marked tool takes the safety arm, words unchanged · G3b the opt-in kind's approval ask keeps its sentence · G4 the question's gate keeps exactly its other members, the row its scope / deadline / opt-in / pending call, the descriptor the same bytes. Baseline (both plain arms as before) red on G1 / G1b ×2 / G2 / G3, green on G1c ×3 and the three controls; green here. Mutations, each red on the cells it should and green elsewhere (per cell, per fixture): every ask worded as a question ⇒ G2 / G3b · the old irreversible sentence restored for the unattended arm's approval ask ⇒ G2 · the opt-in arm wording a question as an approval again ⇒ G1 / G1b ×2 / G3 · shared sentence edited ⇒ G1 / G1b ×2 / G3 · extra gate member ⇒ G4 · card member dropped ⇒ G1 / G1b ×2 / G3 / G4 · safety sentence edited ⇒ G2b · the opt-in arm's judge widened to an ordinary tool ⇒ G3b · opt-in arm filing the unattended kind ⇒ G1 / G1b ×2 / G3 / G3b / G4 · the question decline removed ⇒ G1c ×3 · the refusal sentence worded as an irreversible approval ⇒ G1c ×3 · one word of the approval sentence changed in its one function ⇒ G2 / G3b (both arms read the one mint).
|
|
492
|
+
- **Consumers.** @client-core @cli @test: a card's `reason` bytes change for the asks above — a renderer or black-box fixture that pins an old literal (either pair above) re-pins to the new one; read `gate.kind` / `approvalCard`, never branch on the text (display metadata). @server: question rows (kind `human`) and approval rows on the unattended kind read the new words; the listing and the structural reads (`gateKind`, `approvalCard`) are unchanged.
|
|
493
|
+
|
|
494
|
+
### A request's reasoning resolution is what its thinking keys say — the wire applier says it where it writes them, the resolver and the supported-tier read are readers of that answer, and a leg's seat corrects from its own first request's answer, stamped on that request's terminal message (#4016, #3999 S1; ruled C-R1341 and its note 1; behavior change: `effective` reads `on` / `provider_default` where it echoed a tier, `off` legs get a seat, the correction reads the leg's own request only; narrowing: four, below, open to veto @server (N-2 also @cli @client-core); widening: none; hard-breaking: five, below, open to veto @cli @client-core @server @sdk; closed sets: `ReasoningEffective` +`on`; exports: −`resolveBinary`, −`AppliedReasoning.wireFacts`, +`AppliedReasoning.carried`, +`ReasoningCarriage`; wire: one, below) (@cli @client-core @server @sdk @test)
|
|
495
|
+
|
|
496
|
+
- **Rule.** One computation: the wire applier (`applyReasoningToRequest`), in the arm that writes a request's thinking keys, says what they mean — `carried.said`: a tier the keys name (the explicit `off`; the lane's tier key — the clamped effort value, or on the Messages budget form with no budget knob the budget the tier's share sized; where Messages declares `effortLevels`, `output_config.effort`), `on` (thinking on, no tier named), or `provider_default` (no thinking key). A resolution is `(requested, said, format)` with the entry's `api` as `endpoint`: `effective` = `said`, `graded` ⇔ six enabled tiers, `clamped` ⇔ `effective ≠ requested`, `dropped` ⇔ `provider_default`. The leg's seat reads the leg's own first request's answer off that request's terminal message.
|
|
497
|
+
- **Before.** `resolveReasoning` re-derived the answer from the model entry in its own format table (eleven arms mirrored from the applier) and five shapes disagreed with the wire: an out-of-set dialect name (resolver `high`, wire nothing), Anthropic adaptive with no `effortLevels` (resolver `max`, wire only `{type:"adaptive"}`), Messages `minimal` (resolver `minimal`, wire `low`), an `off` the wire cannot spell (resolver `off`, wire nothing), a withheld effort carrier (resolver echoed the tier, wire nothing). The leg's correction rode an `AsyncLocalStorage` channel any nested request inside the leg could reach: a mod's `$.model.complete` at `turn.start` on the leg's own model with a small cap set the seat to `provider_default` while the leg thought.
|
|
498
|
+
- **Now.** The applier answers; `resolveReasoning(requested, model, facts?)` is the applier's answer for a request of that tier; `reasoningEffortLevelsOf(model)` is the enabled tiers a request carries as an effort value and says as itself (rank order). The stream engine stamps a request's terminal message — success and engine-ended failure — with `{ model, level, carried }` (a `WeakMap` keyed by the message object: no member, no transcript byte); the seat reads every `message_end` of the leg's harness and takes the first message stamped with the leg's own `Model` object. A mod's, a tool's, the classifier's, the compaction summary's or the advisor's request never reaches the leg's `message_end`; a fallback hop or a degrade target is another object. Messages `minimal` reads as `low` inside the clamp set (wire unchanged on 490 set × tier cells).
|
|
499
|
+
- **Hard-breaking — five, three sentences each.**
|
|
500
|
+
- *HB-1 `effective` gains `on` (open to veto @cli @client-core).* What changed: qwen / zai / qwen-chat-template, deepseek and together with no effort, Anthropic adaptive with no `effortLevels`, a Messages budget a knob fixed (after the first request) and openrouter's null-mapped `reasoning: {}` now resolve `effective: "on"`, `graded: false`, `clamped: true` instead of echoing the requested tier. Who is affected: the cli's `effectiveReasoningLine` (`effectiveTurnFactsStore.ts`), client-core's view (`effectiveFacts.ts:21`) and `-p --output-format json`'s `_sema_effective_reasoning`, and any host that feeds `effective` to a tier selector. Compensation: `requested` still names the tier; `clamped: true` lets the clamped note speak; the word is a closed-set member (`ReasoningEffective` stays a literal union).
|
|
501
|
+
- *HB-2 every "nothing reached the wire" is `provider_default` + `dropped: true` (open to veto @cli @server).* What changed: an `off` the wire cannot spell (stock openai or Responses with no off value, Anthropic adaptive, a model that holds no reasoning) and a withheld effort carrier (`supportsReasoningEffort: false`, a `null` levelmap entry on an effort-only wire) now resolve `provider_default` / `dropped: true` (they read `off` or echoed the tier). Who is affected: a consumer that reads `effective` as a tier on such endpoints, and the cli's published known-limit sentence about `_sema_effective_reasoning.effective` reading `off`. Compensation: `requested` is unchanged and `dropped` is the existing discriminant.
|
|
502
|
+
- *HB-3 `graded` / `clamped` follow their definitions (open to veto @client-core @cli).* What changed: `off` is always `graded: false` (it was `true` where the wire spells it), switch-only shapes `graded: false`, and every shape whose `effective` differs from `requested` is `clamped: true` (adaptive, Messages `minimal`, the `on` and `provider_default` shapes were `false`). Who is affected: tests and views that pin the old booleans; client-core's view comments read `graded` as "the endpoint accepts graded tiers" and should read "the wire names an intensity tier". Compensation: the definitions are unchanged — they are now executed literally.
|
|
503
|
+
- *HB-4 `off` legs have a seat (open to veto @cli @client-core @server).* What changed: a leg whose tier resolves to `off` (explicit, a role's, `Model.defaultThinking`) now carries `effectiveReasoning` and a `reasoning.resolved` frame — `off` where the wire spells it, `provider_default` / `dropped` where it cannot (every `off` leg on a model that holds no reasoning). Who is affected: a consumer that reads the seat's presence as "thinking was requested on", and the cli's clamped wording, which reads like a fault for `requested off → effective provider_default` and changes in the same release. Compensation: `requested: "off"` says so; a leg with no tier still has no seat.
|
|
504
|
+
- *HB-5 two root members retired (open to veto @server @sdk).* What changed: `resolveBinary` and `AppliedReasoning.wireFacts` are removed. Who is affected: a host that called them directly (none of the eight repositories' `src` does). Compensation: `applyReasoningToRequest(...).carried` answers the same question for every lane.
|
|
505
|
+
- **Not breaking — the resolvers' model parameter (a wire-manifest reading).** `resolveReasoning(requested, model, facts?)` and `reasoningEffortLevelsOf(model)` take `model` as the named `ReasoningRequestModel` where the parameter was written inline: the same four optional members (`api`, `reasoning`, `compat`, `thinkingLevelMap`) plus an optional `id`, so every earlier argument compiles (a value carrying an `id` that is not a string is the one that no longer matches).
|
|
506
|
+
- **Narrowing (open to veto @server).** N-1: the model door refuses a `compat.thinkingFormat` outside the closed set on any entry (`config.unknown_role` / `not_a_model_ref`, naming the member; one bad entry refuses its catalog, as a bad `defaultThinking` does), and the wire applier, `resolveReasoning`, `thinkingOffSpelled`, `offSpelling`, `lowestReasoningSeat` and `reasoningEffortLevelsOf` throw `config.invalid_argument` (`detail.field` = `compat.thinkingFormat`) where each answered its own way; `null` and absence stay `openai`. N-2: `applyReasoningToRequest` refuses a `level` outside the seven tiers (`detail.field` = `applyReasoningToRequest.level`) where it wrote nothing; a host that calls a brain directly with such a value (a typo, a legacy boolean) gets no request sent and a stream that ends on an error message naming the member (no code — the Runner judges the tier at its own seats first), where the request went out with no thinking key — the user face (open to veto @server @cli @client-core): a request that carries a legacy boolean or a garbage `reasoning` value to a brain used to be sent, with no thinking, and is now refused with nothing sent. N-3: `reasoningEffortLevelsOf` no longer lists `minimal` for a Messages entry that declares it (the wire always sent `low`), and lists in rank order. N-4: `applyReasoningToRequest` refuses a body that already carries one of its lane's thinking members (completions `reasoning_effort`, `thinking`, the whole `reasoning` object — openrouter reads its presence as the switch, so a host `{}` or `{ exclude: true }` would turn reasoning on while the request said `provider_default` — `enable_thinking`, `chat_template_kwargs.enable_thinking`; Responses `reasoning.effort`; Messages `thinking`, `output_config.effort`, a `clear_thinking_*` edit), whatever the tier, `undefined` included, and a member it writes into that is not an object; its Messages budget-form refusals (no `max_tokens`; a cap or knob that does not read as a finite number) carry `config.invalid_argument` with `detail.field` = the input, where they were bare `TypeError`s.
|
|
507
|
+
- **Wire — one change, host bodies only (open to veto @server).** A host body's other members of the objects the applier writes into are kept and the applier's merged in, where the applier overwrote the object whole: `output_config.format` beside `effort`, `context_management.edits` = the thinking edit then the host's own (CC 316 order), Responses `reasoning.summary`. The engine's brains send none of these members: their bytes are unchanged, including Messages budgets (shares by the requested tier) and openrouter's `reasoning: {}`.
|
|
508
|
+
- **Not changed.** The seven tiers and their clamp; the levelmap; budget shares; `endpoint` (the entry's `api` label); the degradation law (the seat is the primary model's at leg entry); `TaskResult` keys from this section (the release adds `limitCrossed`, #4184, and `refusedAsks`, #4111, below); HTTP and server wire keys.
|
|
509
|
+
- **Rules.** Retired: the resolver's format dispatch (`dispatchReasoning`, eleven arms) and its helpers (`droppedResolution`, `BINARY_FORMATS`, `completionsEffortHonored`, `resolveBinary`); the Messages `minimal → low` write-site special case; the seat's two `!== "off"` guards and its hand-listed six-key dedupe (now whole-structure); the "no dialect" reading of a bad name and the third reading of the supported tiers; four whole-object overwrites (H5); the reasoning ALS channel (store, leg scope, liveness bit, the third sink and the two no-op reasoning shields). Added: `said` in each lane's write arm (one `saidOf` law), `resolutionOf`, the word `on`, one model-door cell, one dialect judge, one owned-member table with one merge, one engine stamp point and one seat read. Special cases: none (an unstamped or other-object message is not "the first" — the identity sentence read as written).
|
|
510
|
+
- **Known limits.** KL-3999-a (adjacent tiers share one budget below `max_tokens` 4099 / 2562 on the budget form), KL-3999-b (a deployment Brain that copies the `Model` object never corrects), KL-3999-c (a terminal minted outside the engine carries no stamp: the seat keeps the entry value), KL-3999-d (openrouter's `{}` read as `on`, unverified live).
|
|
511
|
+
- **Pins.** `test/backlog3999-reasoning-single-source.test.ts` G1–G20 (G1 315-cell matrix resolver = applier = key reading; G5 / G13 / G19 / G20 real Runner legs over real brains; G13a a real mod at `turn.start`; G17c each engine brain ends each request on its own terminal object — the premise the object identity rests on). The reasoning wire matrix (`backlog2163` E1) keeps its golden: every cell the engine still sends is byte-identical, and the cells it now refuses (a tier outside the seven, a dialect name outside the set) keep their golden rows and are pinned as sending no request. Re-pinned: backlog2232, backlog2163, backlog2395, backlog2566, backlog3998, reasoning-intensity, reasoning-levelmap-resolver, backlog367, observability, open-responses, openai-thinking-gate-contract and the import sites of the moved resolver functions; `test/anthropic.test.ts` RB-482 #11 (a legacy boolean and a garbage tier: no request sent, the error names `applyReasoningToRequest.level`; it read as absence — re-pinned by a stage fix).
|
|
512
|
+
- **Consumers.** @cli: `effectiveReasoningLine` prints `on` and the `off → provider_default` row; change the clamped sentence for `requested off`. @client-core: `effectiveFacts.ts:19-24` view comments (`graded`, `effective`'s words, `dropped`) and the `-p` JSON field gain `on`. @server: the capability face's `reasoningEffortLevelsOf` / `thinkingOffSpelled` / `lowestReasoningSeat` throw on an out-of-set dialect name (you filter before core); the result-text leaf keeps literal unions. @sdk: two retired members in the d.ts. @test: G2 / G3 / G6 / G13a are wire-observable.
|
|
513
|
+
|
|
514
|
+
### A person is asked only about a plan the card in front of them holds — a plan-review request whose call ended in error, or whose plan card did not ride its `tool_end`, no longer parks; `ExitPlanMode` refuses a plan its card cannot carry before it asks (#4120; behavior change: narrowing — two forms that parked now continue, open to veto, D-4120-a; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — no key, no word, no frame member) (@server @cli @client-core @test)
|
|
515
|
+
|
|
516
|
+
- **Rule.** A plan-review pause asks a person to approve the plan the requesting call's `tool_end` CARRIED. A request whose call's end showed nothing — an error end (the tool failed, or a station withheld the result after the body ran) or a result holding a plan card the frame did not carry (over the card bound, or with no `plan` string) — is withdrawn at that call's end, the one place the frame's card is recorded (`notePresentedPlan`); the park mint refuses it in its one `plan_review park skipped: …` sentence; the turn boundary gains no condition. A batch's later requests queue behind the first (they used to be dropped): a withdrawn first request yields to the first one not withdrawn, so when two plan calls share a batch and the first is withheld, the second — whose call was told "The run is pausing for approval" — is the one put to the person. The first-party `ExitPlanMode`, both arms, measures its card exactly as the frame will (`MAX_TOOL_CARD_CHARS` / `toolCardChars`, `src/core/tool-card-bound.ts`) and refuses a card over the bound BEFORE it requests review.
|
|
517
|
+
- **Before.** Both forms parked on `plan_review` (`needs_review`): (1) a `tool.call` extension failing after the body (a throw or a `{ deny }`; since #4143 in this release, below, a throw after the body is skipped and the result stands, so form (1) is now an explicit `{ deny }` only) — `tool_end` `isError: true`, no card; (2) a plan whose card JSON passes 65 536 characters — the card left off the frame while the model read "The run is pausing for approval". A person was asked to approve a plan no card carried (measured on the base: `paused`, gate `plan_review`, `pendingAction` `{ kind: "plan_review", requestedByCallId: "plan-call-0" }`).
|
|
518
|
+
- **Now.** (1) — a `{ deny }` after the body — continues: the model reads the withheld result, no row is written (an extension that throws after the body is skipped since #4143: the card rides and the run pauses for that plan, pinned in `test/backlog4120-plan-card-withheld.test.ts`), and `onError` reads `plan_review park skipped: the call that requested it ended in error (its result failed or was withheld), so no plan card reached a reviewer (the plan-review request was dropped; the run continues without pausing)`. (2) is refused at the tool — no request, no row, plan mode not exited: `The plan was NOT submitted for approval and plan mode was not exited: its approval card is N characters as JSON, over the 65536 that the card can carry. A quote, backslash, tab or line break counts as two characters, any other control character as six, and a character outside the Basic Multilingual Plane (most emoji) as two. Shorten the plan and call ExitPlanMode again.` (one sentence for both arms). A caller tool's `ctx.requestReview` reads the same way: an error end, or a plan card over the bound, ⇒ no pause (`… did not ride that call's tool_end …`); a success with no plan card still parks.
|
|
519
|
+
- **Narrowing (open to veto @server @cli @client-core; D-4120-a).** Runs that ended `paused` / `needs_review` in the two forms now go on to their next terminal (`completed` in the pins). D-4120-a: the reference refuses an over-size plan on its diskless arm only (32 768 bytes of input JSON, before its approval card is shown); this engine refuses on both arms at its frame's card bound (65 536 characters of card JSON), because both arms put the plan on the card the frame carries. The not-pausing arm (a per-task env with no suspend target) is withdrawn by the same rule, and its error face still says the suspend-target gap — the mint says that reason first.
|
|
520
|
+
- **Three questions.** Who needs it: the person at an approval surface (never asked to approve a plan they cannot see), and the model (told to shorten the plan instead of being told it is pausing). Who is hurt: a host or test that relied on the pause in either form. Compensation: the model reads why and can call again; the deployment's error face names every withdrawal; a carried card parks exactly as before.
|
|
521
|
+
- **Not changed.** The turn boundary's honoring conditions; `planContentHash` / `requestedByCallId` (a carried card parks with its digest); the not-pausing arm's sentences; the card bound (65 536) and the text cap (16 384) — only where they are spelled; how a `tool.call` extension's failure after the body is answered.
|
|
522
|
+
- **Rules.** Card-bound spellings 2 → 1 (the projection's inline `* 4` and its own JSON measure now read `tool-card-bound.ts`, which the tool reads too); plan-review skip sentences 3 inline → 1 function (`parkSkipped`); the end settlement gains one outcome (`withdrawn`) beside `presented`, at the one site that already read the frame's card, and the cell one queue (`later`) so "the first request wins" reads "the first request that may still show a plan wins"; the turn boundary gains no condition (it clears the queue with the request).
|
|
523
|
+
- **Pins.** `test/backlog4120-plan-card-withheld.test.ts` — W1 form (1) on a real Runner · W2 form (2) on a real Runner · W3 the bound both sides on a real Runner · W4 both arms over / at the bound · W5 the not-pausing arm (keep cell: its gap sentence) · W6 caller tools (over-bound card, error end, and the no-card control) · W7 the settlement table · W8 one bound, one measure · W9 two plan calls in one batch, the first withheld: the second parks; base 10 red / 2 green before W9 (W5 and W6's control are keep cells). `test/backlog4103-plan-content-hash.test.ts` ⑩ withheld / Q4 / Q4b re-pinned (paused → completed). Mutations M1–M10 and #4103's M16 each red.
|
|
524
|
+
- **Consumers.** @server: a plan-exit result your host's extension withholds no longer parks — read the `onError` sentence. @cli @client-core: no wire change; a plan over the card bound arrives as an error `tool_end` carrying the refusal sentence, never as a pause whose card is missing. @test: real-binary steps for the reference's two forms are with the main session.
|
|
525
|
+
|
|
526
|
+
### A plan-review pause carries the digest of the plan text shown for review — `planContentHash` on the paused cause, the parked row and the inbox row; and the package mints and reads its `sha256:` text form in one place (#4103, #4119; behavior change: additive — three optional members; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: `PendingAction` `plan_review.planContentHash?`, `CheckpointSummary.planContentHash?`, `TerminalCause` `paused.planContentHash?`) (@server @client-core @test)
|
|
527
|
+
|
|
528
|
+
- **Rule.** The text shown at a plan pause is the `plan` string on the `exit-plan-mode` card the requesting call's `tool_end` CARRIED — `structured.plan`, the projection of the final result after every post-execution station; a card the frame leaves off (over its size bound) showed nothing a reader holds. The end projection records it from the card it just put on the frame (`notePresentedPlan`, only for the call the request names, only once); the park mint digests it once — `sha256:` + the lowercase hex SHA-256 of its UTF-8 bytes, unnormalized, untruncated (71 characters) — onto the row, and the paused cause and the summary echo the committed row through one reader (`planReviewContentHashOf`, beside `planReviewRequesterOf`). The digest says what was SHOWN, not who asked: a caller tool that returns that card carries one too.
|
|
529
|
+
- **Before.** A card could join a plan pause to the call that presented it (`requestedByCallId`), but nothing said which text the person was asked to approve: a client holding the plan string from a frame could not tell whether it was the one parked. The file arm of `ExitPlanMode` also parked without `requestedByCallId`, though its call had an id (both arms now bind through one function, `reviewRequestFor`, which a caller tool's `ctx.requestReview` uses too).
|
|
530
|
+
- **Now.** A `plan_review` row, its paused cause and its inbox row carry `planContentHash` when the requesting call's frame carried a plan card; absent (key omitted) otherwise — an error, a result an extension withheld, a card left off the frame for its size, a requester that returned no card — and on rows a deployment minted itself or an earlier engine wrote. Absent means "nothing to compare", never "an empty plan". Written once; no decision, edit or reopen rewrites it. It is not the gate's identity (two pauses showing the same text carry equal values under different `checkpointId`s) and no resume reads it. To use it: digest the plan string you received, before any display normalization, and compare; do not digest the plan file on disk (its reader drops a BOM) or the model-facing fenced text. The member is SHOWN in the row's requirement table (ignorable): a row with it reads on an engine without it, and the other way round. Unsalted: it stays off traces, notices and error sentences.
|
|
531
|
+
- **One `sha256:` text form.** `textContentHashOf` / `isTextContentHash` (`src/core/canonical-json.ts`) replace eight inline copies of `"sha256:" + hex(sha256(string))` (the prompt manifest's unsalted section digest, a locale row's `sourceDigest`, a prompt artifact's digest, the turn snapshot's digests, the prompt epoch's `digestOf`, the git status frame's hash, both rule-store checksums) and three copies of its shape check; every value is byte-identical to what each copy minted (how each caller builds the string it digests is unchanged).
|
|
532
|
+
- **Not changed.** The plan text stays off the gate; `requestedByCallId` keeps its reading (now also present on the file arm); the plan-review resume, its decision vocabulary and `editedPlan`; what a run whose `ExitPlanMode` result is withheld, or whose plan card the frame leaves off for its size, does is #4120's (this release, above): the withheld one's request is withdrawn, the oversized plan is refused at the tool, and nothing parks — this section's mint is never reached for either.
|
|
533
|
+
- **Rules.** Digest mints of the text form: 8 → 1; shape readings: the three copies → 1 (a fourth is the new reader, which reads through it) — one more spelling of the shape stays, the prompt artifact store's file-name match (`src/prompt-assembly/artifact-store.ts`, which takes the hex out), so the package holds two; the requester binding: 3 sites (parameter arm, file arm, caller wrapper) → 1 function, and the review cell's own re-filter of the request retired (one guard on one path). Added: one member on three faces under one name, one reader, one recording point (the end projection, which already reads the same `details` for the card).
|
|
534
|
+
- **Pins.** `test/backlog4103-plan-content-hash.test.ts` — ① parameter arm, every face (CRLF and trailing blanks kept) · ①b a PreToolUse argument rewrite: the card decides · ② file arm with a BOM, the id present · ③ a caller tool without a card (forged options) / ③b with one · ④ first request wins · ⑤ the reader · ⑥ SHOWN, ⑥b readers both ways · ⑦ written once (edit, env_failed reopen, re-shown verbatim) · ⑧ the static scan, the one function against the old formula, every routed site, a rule bucket written with the old formula · ⑨ a lone surrogate · ⑩ the final result (a withheld result, the end projection's arms) · Q4 a 70 000-character plan whose card the frame leaves off carries no digest · Q4b the card bound from both sides · ⑪ off traces, notices and errors. 16 of 23 cells red on the base; the controls (③, ⑥b, ⑧ sites, ⑧ bucket, ⑧ manifest, ⑩ withheld, Q4) green on both.
|
|
535
|
+
- **Consumers.** @client-core (CC-1497): digest the `structured.plan` string you received from the requesting call's `tool_end` (before CRLF normalization) and compare with `planContentHash` on the paused cause or the inbox row; treat absence as "cannot say", not "different". @server: pass the three members through verbatim (the pending-queue row and the run's paused cause); do not recompute, rewrite or log them. @test: one black-box cell per arm — the three faces agree with your own digest of the card's plan.
|
|
536
|
+
|
|
537
|
+
### A `tool.call` extension that fails after the tool ran is skipped and the tool's result stands; only an explicit `{ deny }` withholds a result that ran, and the model is told who withheld it, why, and that the call ran (#4143; ruled C-R1427; behavior change; widening: one, below, open to veto — a safety-shaped one @server @cli @client-core @test; hard-breaking: the mods contract 10 → 11, below; closed sets: none — `MOD_EVENT_TABLE["tool.call"].station.failure` moves within `MOD_FAILURE_ANSWERS`; exports: none added, none removed; wire: the version literal `KERNEL_CONTRACT_VERSION` 10 → 11, model-facing sentences) (@server @cli @client-core @sdk @test)
|
|
538
|
+
|
|
539
|
+
- **Rule.** `tool.call` reads the kernel's go-ahead failure column `refuse_until_run`, the one `agent.spawn`, `session.send` and `session.compact` already read: a handler that fails while the tool's body has not started refuses the call (unchanged); a handler that fails once its own `next(e)` ran the body — it threw, ran past its budget, or answered a form the site does not take — is skipped, and the result that `next` answered stands, said on `onError`. A `{ deny: <reason> }` over what `next` answered once the body ran is the one answer that withholds a result that ran.
|
|
540
|
+
- **Before → now, what the model reads.** (1) `next`, then a throw, an overrun, an answer that is not data, or a `{ result }`: before — an error, `a tool.call extension crashed (extension "x") after this call to "T" ran: a hook that checks this result could not hand back its answer, so the result is withheld — the call ran (fail-closed)`; now — the tool's own result, as the tool gave it, and `onError` reads `x: tool.call: <reason> — skipped; its last next() run's result stands` (was `… — failed closed: the dispatch ends here`). (2) `next`, then `{ deny: "why" }`: before — the same withheld sentence as (1), the reason never shown, `onError` `x: tool.call: a deny after the tool started (once the tool runs, nothing answers in its place) — failed closed: …`; now — an error, `this call to "T" ran, and extension "x" withheld its result: why — the call was executed and whatever it did has happened; it is only its result that is not shown` (the reason through the engine's extension-text pass, bounded at 300 characters), nothing on `onError` (an answer, not a failure). (3) A throw before `next`: unchanged — refused, `the call was NOT executed (fail-closed)`.
|
|
541
|
+
- **Widening (open to veto; safety-shaped, recorded under C-R1400 ③; @server @cli @client-core @test).** A result that a crashing, overrunning or malformed `tool.call` checker used to keep from the model now reaches it. *Who needs it:* the model and the person — a checker's own bug no longer erases the result of a call that ran (a write that happened, read back as "withheld", invites a second write), and extension authors get the reference's law. *Who is hurt:* a deployment that relied on "the checker crashed, so the result is withheld" as its line of defence — a redacting checker that throws on content it cannot scan now lets that content through; a `{ result }` given after the body (a rewrite) is skipped too, and the tool's own text reaches the model. *Compensation:* the deny form — a checker that means to withhold answers `{ deny }`, or catches its own failure and denies; every skip is said on `onError`, naming the extension and its reason.
|
|
542
|
+
- **Hard-breaking (three sentences).** The mods kernel contract is 11 (`KERNEL_CONTRACT_VERSION` 10 → 11, in the core root and in `@sema-agent/mods-kernel`): a package whose `hooks/hooks.json` states `"contract": 10` is refused `config.mods_invalid` naming both numbers, so a package is rebuilt against this engine and states 11. `MOD_EVENT_TABLE["tool.call"].station.failure` is `refuse_until_run` (was `refuse`), so a host that reads the column reads the go-ahead law. A handler that failed after the body used to have the result withheld and now has it kept, and a handler that denied after the body used to break the site's contract and is now taken.
|
|
543
|
+
- **The reference (Claude Code 2.1.321).** A throw after `next` is skipped and the result stands (`ha`, "skipped; its last next() run's result stands", cli321.js:278661 — measured on the real binary by test); a deny after `next` is the withholding ("ran, and a plugin withheld its result: ", cli321.js:288437 — measured); an overrun takes the same skip branch (cli321.js:278661, read). Its one withholding without a deny is a failure to carry an answer across its VM boundary ("a hook that checks this result could not hand back its answer", cli321.js:279676, read); this engine runs extensions in-process, so that sentence stays only where it is true here — the `classic.PostToolUse` seat that could not hand a result to its handlers. Stricter than the reference on two forms, kept by ruling (C-R1427 note 1; divergences `mod-tool-call-late-answer-violation` and `mod-hook-failure-fails-closed`): a `{ result }` in the tool's own shape after `next` is taken there as the tool's result ("{ result }", cli321.js:278686; measured on the real binary by test), here the handler is skipped and the tool's own result stands — a rewrite of a result that ran has its seat at `classic.PostToolUse` (`updatedOutput`); a throw or a hang before `next` is skipped there and the tool runs in its place (`xa`, "skipped; what is below it ran in its place", cli321.js:278661; measured by test), here the call is refused, as at the other go-ahead stations and the judgment seats.
|
|
544
|
+
- **What a deny is.** An answer record that carries its own `deny` key (`Object.hasOwn`; a key present with the value `undefined` counts), whatever sits beside it — `{ deny, result }` is a deny, the result beside it not taken. Its reason is the key's value when that is a non-empty string; with any other value (`""`, a number, `null`, an object, `undefined`) it is still a deny — refused before the body, withheld after it — the model reads `(it gave no readable reason)` where the reason would be, and `onError` reads `<name>: tool.call: its deny gave no readable reason (a reason is a non-empty string) — the deny stands`. Before, only the exact one-key, non-empty-string form was a deny: the rest broke the site's contract, which refused before the body (unchanged direction, other sentence) and withheld after it — under this release's column that would have kept the result, the opposite of what the extension said. An answer the bus cannot receive as data (a function or an accessor in it) is still not read at all. *Rules:* one reading of "a deny" for the answer check, the ranks and the station; none added. *Seat:* `ToolCallStationInput.report` (the leg's report sink, as `session.compact` and the row and receive seats already take it).
|
|
545
|
+
- **Not changed.** A failure before the body (the refusal, its sentence, `notExecuted`); a well-formed deny before the body; a `{ result }` in place of the body under its two facts; a body that threw keeps its throw whatever a handler answers over it (a deny over a thrown body is still a violation, now skipped, the throw standing); a cancel; the `classic.PostToolUse` seat and its sentence; a frame the withheld result carried (a ReadInbox message) is withheld with it.
|
|
546
|
+
- **Rules.** `tool.call` failure readings 2 → 1: the station's own "failed after the body" arm is retired — the kernel's column answers it — and with it `failedHandlerName` (its one caller). The station's answer check keeps its rule count: a deny over what `next` answered changes reading once the tool ran (taken, a withholding) and stays a violation before it.
|
|
547
|
+
- **Contract.** `KERNEL_CONTRACT_VERSION` 10 → 11, this window's one raise; entry 11 names the column and the deny.
|
|
548
|
+
- **Pins.** `test/backlog4143-tool-call-throw-keeps-result.test.ts` — T1 `next` then a throw keeps the result and `onError` says so · T2 `next` then a deny withholds, the sentence carries the reason and says the call ran · T3 a throw before `next` is refused (control) · T4 an overrun · T5 an answer that is not data · T6 a `{ result }` after the body — each keeps the result · T7 an inner extension's deny handed up names the inner one · T8 the reason is bounded · T9 a deny over a thrown body withholds nothing, the throw stands · T10 a deny is its key: five unreadable reasons after the body withhold, `{ deny, result }` withholds with its reason, and before the body both refuse. Re-pinned: `test/backlog2675-withheld-after-run.test.ts` ① ①b ⑥ ⑦ (⑦'s fixture now denies); `test/mods-tool-call.test.ts` C2 C3 C7 L5; `test/backlog4120-plan-card-withheld.test.ts` W1 (the deny form, and a throw after the body: the card rides and the run pauses) and W9 (the deny form); `test/backlog4103-plan-content-hash.test.ts` ⑩ (the deny form); `test/mods-kernel-wave2.test.ts` and `test/backlog2460-clock-every.test.ts` (contract 11); the three packages under `test/fixtures/mods/` state 11; `test/backlog3058-agent-spawn-entry.test.ts` S8 (a throw after the delegation ran keeps the result and is said on `onError`, the deny form withholds it; `$.agent.spawn` answers `{ model, agentId }` on both — re-pinned by a stage fix).
|
|
549
|
+
- **Consumers.** @server: every mods package you ship states contract 11; a `tool.call` checker of yours that relied on crashing to withhold must answer `{ deny }`; the model-facing sentences above changed. @cli @client-core: a `tool_end` that is `isError` with no `notExecuted` after a deny now carries the extension's reason in its text; a checker's crash shows as the tool's normal result plus an `onError` line. @sdk: the contract number moves, and `MOD_EVENT_TABLE["tool.call"].station.failure` reads `refuse_until_run`. @test: the two forms read but not measured on the reference (a throw before `next`; a `{ result }` after `next`) are yours to measure.
|
|
550
|
+
|
|
551
|
+
### A session's tree stands for one checkout of its repository — roots standing in two linked work trees of one isolated repository are refused `workspace.isolation_unavailable`, naming both roots and both checkout tops, where the second was placed in the first's tree (#4009; design/2378 r1 S2b follow-up; behavior change: a leg whose declared roots stand in two different checkouts of one isolated repository is refused before any git write, where it ran with the second root mapped into the tree of the first; narrowing: three, below, open to veto @cli @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @server @client-core @web-client @test)
|
|
552
|
+
|
|
553
|
+
- **Rule.** A repository's linked work trees are other checkouts (another branch, other files), and a session's tree is one tree started from one commit. Every declared root of the leg (the task root and both directory lists, listed in `workspace.roots` or not) that stands in an isolated repository, outside the session's own trees, must stand in ONE checkout of it (the top of the work tree it is in, `git rev-parse --show-toplevel`); two ⇒ `workspace.isolation_unavailable`, the sentence naming both roots and both checkout tops, judged before any git write on the first leg (no tree, no record, no frame) and again on every later leg (the record and the tree stay as they were).
|
|
554
|
+
- **Before.** A root in a second work tree of the repository was placed at its path below that work tree's top inside the tree minted for the first (the main checkout `r` and a linked work tree `r2` on branch `other`, both listed: `r2` was reported `at` that tree's root), so a leg that declared the `other` checkout read and wrote a tree started from one commit (the default branch tip, or the first listed root's HEAD under `baseRef: "head"`), and nothing said so.
|
|
555
|
+
- **Now.** The leg is refused; the declaration is fixed by naming the roots of one checkout, or by working without isolation. Two directories of one work tree (the main checkout, or one linked work tree) still share one tree; two repositories each in one work tree still get two trees; a root inside the session's own tree is that tree's, not a second checkout; a nested repository or a submodule is its own repository, judged by itself.
|
|
556
|
+
- **Narrowing (open to veto @cli @server).** A declaration that lists the task root and a root in another work tree of the same repository is now refused on the first leg (`workspace.isolation_unavailable`, no tree minted) where it ran with the second root standing in the first's tree.
|
|
557
|
+
- **Narrowing (open to veto @cli @server).** The same two checkouts reached through an unlisted root (an `additionalDirectories` or `additionalReadDirectories` entry in a second work tree of an isolated repository) are refused too, where the entry followed the tree.
|
|
558
|
+
- **Narrowing (open to veto @cli @server).** A later leg that adds such a root to a session whose tree is already minted is refused with the record and the tree untouched, and the next leg without that root runs.
|
|
559
|
+
- **Not changed.** The record's shape, the frame, the error code set (the code is the existing one), the capability judge, the release and orphan doors; one tree per repository (no tree per work tree: that would be a second tree and a second base for one repository, the reverse of the one-tree rule).
|
|
560
|
+
- **Rules.** Added: one judge (`refuseSplitCheckout`), read at the two seats the exposed-base judge is read at. No special case is added: it reads the mapping's own placement (`rootPlaceOf`), and the mapping is unchanged.
|
|
561
|
+
- **Pins.** `test/backlog4009-two-worktrees-one-repo.test.ts` — G1 the task root and a listed root in two work trees ⇒ refused naming both roots and both tops; no tree, no branch, no record, no frame, no write (also reversed, and for directories inside the work trees) · G2 two directories of one work tree share one tree (the main checkout; a linked work tree started at its HEAD) · G3 two repositories each in one work tree ⇒ two trees; an unnamed second work tree is no root · G4 the same split through an unlisted write root and an unlisted read root · G5 a later leg adding the root is refused, the record and tree stay, the next leg runs · G6 a later leg's root inside the session's own tree is no second checkout · G7 a root reached through a symlink is judged where it lands (a link to the other work tree is refused with its real top, a link to the task root's own checkout is one checkout) · G8 a work tree made inside the main checkout's directory is its own checkout · the judge is one function read at both seats, before the first mint.
|
|
562
|
+
- **Consumers.** @cli (L-2145, multi-root tree rendering): a refused declaration is a failed leg with no `workspace_changed` frame — render the refusal, not a tree. @server (S-1950): forward `workspace` unchanged; a leg refused `workspace.isolation_unavailable` here is not a store fault — its sentence names the two roots to change. @client-core / @web-client: a per-folder isolation switch must not offer two work trees of one repository together. @test (black box, wire-observable): two work trees of one repository as the task root and a listed or unlisted additional root ⇒ a failed leg, `workspace.isolation_unavailable` naming both, no `.sema/worktrees` entry and no `sema/*` branch left, the other work tree untouched.
|
|
563
|
+
|
|
564
|
+
### A session over several folders works in one tree per repository — `TaskSpec.workspace.roots` lists the roots whose repositories are isolated, every declared root passes through one mapping into its repository's tree, and the frame says where each additional root stands (`workspace_changed.directories`) (#2938; design/2378 r1 S2b, ruled C-R961 K-1–K-6; behavior change: an isolated session's additional roots inside an isolated repository stand in its tree; narrowing: one, below, open to veto @server @cli; widening: one, below, open to veto @server @cli; hard-breaking: none; closed sets: none; exports: none; wire: `TaskSpec.workspace.roots?`, `workspace_changed.directories?`, `workspace_changed.declaration.roots?`, the `session_checkout` entry's `declared.roots?`, additive) (@cli @server @client-core @web-client @test)
|
|
565
|
+
|
|
566
|
+
- **Rule.** The unit of isolation is the repository (its git common directory). The declaration lists roots (`roots`: `"taskRoot"` and / or `{ directory }` naming an `additionalDirectories` entry; absent ⇒ the task root, and `["taskRoot"]` is held as absent); each repository a listed root is in gets ONE tree, `<repository root>/.sema/worktrees/<name>` on `sema/<name>` (one name, each repository its own namespace), minted together on the first leg — all or none. Every declared root — the task root, each `additionalDirectories` and `additionalReadDirectories` entry, listed or not — passes through one mapping: a root whose repository has a tree stands at its place in that tree (its path below its own checkout's top), any other root where it was declared. The leg reads the mapped roots only: the spec's two lists are rebound once, right after the checkout judge, so the advertised-writable fence, the file tools' roots, the `# Environment` block and a delegated child's inherited roots all read the tree's spelling. A writable root in no tree must be shown to stand outside every isolated base checkout: one that holds such a checkout (an ancestor directory, a non-repository task root above a listed repository), or lies inside one without a repository of its own being read there, is refused `config.session_workspace_invalid`, judged on the first leg before any git write and on every later leg. A root placed in a tree must resolve inside that tree (a link committed in the tree's start commit could lead it back to the base checkout); one that does not is refused the same way, before the record is written.
|
|
567
|
+
- **Before.** One tree, of the task root's repository only. A writable `additionalDirectories` entry anywhere in that repository (the checkout itself, a directory inside it, an ancestor) was refused; an `additionalReadDirectories` entry there read the base checkout; a second repository could not be isolated.
|
|
568
|
+
- **Now.** `roots` names the repositories to isolate. Two roots of one repository share its tree (one record cell, one branch) when they stand in one checkout of it — roots standing in two linked work trees of one isolated repository are refused `workspace.isolation_unavailable` (#4009, above); a nested repository or a submodule is its own repository (unlisted it stands in place, listed it gets its own tree); a root inside another tree of the repository is placed by that tree's top. The first leg judges every listed root (the capability judge, its repository, a `{ directory }` must be an `additionalDirectories` entry in the same spelling), every source and the exposed-base rule before any git write; then it mints every new tree, then marks each; a refusal on the way removes, while still clean, each tree of this attempt whose allocation mark is proved absent (a mint that fails leaves none of the trees minted before it; a directory git itself left for the mint that failed is not removed — it is not proved this attempt's — and a retry is refused `name_in_use` until it is gone, as before) and keeps and names the rest; a first leg that crashed before its record takes all its marked trees back under their one allocation. A later leg must keep each listed root it declares in an isolated repository (else `config.session_workspace_changed`); a root it adds stands in its repository's tree when that repository is isolated, in place otherwise. A tree whose repository has a declared source starts at the source's commit, every other tree at `baseRef`; a `baseRef` that starts no tree is refused (`config.session_workspace_invalid`, first leg, before any git write — the shape judge no longer refuses "a source for the task root + `baseRef`" by itself). The `.sema` write-protection row reads every tree a write root is in as that root's checkout (a write root in a tree, never one containing trees). `workspace_changed.directories` lists the leg's additional roots on every leg that has any (`{ declared, at, access, tree? }`, `at` canonical, `tree` present iff the root stands in a session tree; an entry that cannot be resolved is not a root and is not listed). Release is unchanged: one list, each tree judged by itself, without `--force`. The orphan doors call a tree held only when its session's record holds its allocation AND names the tree: a first leg retried after a crash that no longer lists a repository records the allocation without the tree its crashed attempt minted there, and that tree is now listed (`record_elsewhere`) and releasable instead of held by no cell.
|
|
569
|
+
- **Widening (open to veto @server @cli).** An isolated session's writable additional directory that IS the isolated checkout or lies inside it was refused (`config.session_workspace_invalid`); it now stands at its place in the session's tree and the leg runs.
|
|
570
|
+
- **Narrowing (open to veto @server @cli).** An isolated session's `additionalReadDirectories` entry inside the isolated repository now reads the session's tree, not the base checkout (the base checkout's files are no longer readable through it); a delegated child of such a session inherits the tree's spelling. An entry whose place does not exist in the tree — a directory git ignores (`build/`, `node_modules/`), present in the base checkout and absent from a fresh tree — is then no root of the leg: skipped and traced (`config.additional_directory_skipped`) as any entry that cannot be resolved, and absent from `directories`.
|
|
571
|
+
- **Mixed version.** A session whose declaration lists `roots` is not continued by an engine at 7.67.0 or earlier (it refuses the member, or refuses the second repository's root as reaching into a base checkout) — refused, never run in a base checkout; roll the workers that share a session store before a host sends `roots`. A declaration without `roots` digests and records exactly as before (a tree an earlier engine minted is still this declaration's).
|
|
572
|
+
- **Not changed.** The record's shape (`checkouts` was a list from the start; a cell's `root` is the first listed root in its repository, or its source's root), the release door and its answer, the orphan doors' answer for a tree its record names (a tree the record does not name — one a retried first leg minted and no longer lists — now reads `record_elsewhere` and is released by them, pin I5), the capability judge, `null` legs, forks (no inheritance), EnterWorktree (it moves the task root's write location only). The instruction layer (which `CLAUDE.md` files an additional root contributes) is not touched here; the mapping is what it reads next.
|
|
573
|
+
- **Rules.** Retired: the task-root-only placement (`placeInTree`) and the later leg's task-root repository check (one mapping and one listed-root check now answer every root); the base-anchored overlap refusal of additional directories (one judgment: a writable root in no tree holding an isolated checkout); the shape judge's source + `baseRef` rule (one allocation rule). Added: the repository reader (git common directory), the mapping, the spec's one rebinding. No special case is added: one tree and N trees run the same code, and "all or none" is one undo law over the list, not a partial state per tree.
|
|
574
|
+
- **Pins.** `test/backlog2938-multi-root-worktrees.test.ts` — I1 two listed roots of one repository share one tree · I1 / K-4 an unlisted task root and a read-only root follow the tree · I2 / I8 / I14 the environment block, the write fence and a delegated child read the tree; a write in the session's own tree meets no question · I3 a nested repository stands in place unlisted and gets its own tree listed; a root in another tree maps by that tree's top · I4 the second of three mints fails: refused naming that root, no record, the first tree and its branch gone, no third tree · I5 a record that could not be written keeps both marked trees; the next leg takes both back under one allocation; another session's same name is `name_in_use` · I6 one tree gone ⇒ `workspace.checkout_missing` naming it, the other unused · I7 later legs (changed roots, an added root in and out of an isolated repository, an ancestor) · I9 a non-repository task root above a listed repository, an ancestor, a listed directory that is no repository, a listed directory that is no entry, a root inside `.git` (its repository unread) · R-d a link in the tree's commit leading a mapped root back to the base checkout is refused before the record · I5 (retry) a tree the retried allocation does not name is listed and released by the orphan doors · I10 `directories` on a leg with no workspace and across two legs · I11 release one list, clean removed and dirty kept, twice · baseRef (no tree started ⇒ refused; with a second tree it starts that one) · the declaration (`roots` grammar, the default held absent) · the record's normalization keeps `roots` · I15 the write views (a root in a tree, a root containing trees, a prefix sibling, a case alias) · I16 a leg that declares no workspace on a session with no checkout reads no task root and answers on the tick of its record read (the prepare path that refuses a run right after keeps its depth; an additional root costs exactly its own place read) · the one rebinding (once, after the judge, before every reader). Re-pinned to the new shape (K-4): `test/backlog2292-session-checkout.test.ts` (the checkout itself and a directory inside it stand in the tree; an ancestor is still refused; the source + `baseRef` cell moved to the leg-level pin above) and two cells of `test/backlog2292-session-checkout-ownership.test.ts` (the same on a later leg, and with the task root spelled as the tree); `test/backlog3462-orphan-tree-doors.test.ts` T13 intercepts the task root's place read by its new command (`git rev-parse --show-toplevel --git-common-dir`), same refusal.
|
|
575
|
+
- **Consumers.** @cli (L-2145, multi-root tree rendering): render one row per `session` entry and place each additional root by `directories[].tree` / `at`. @server (S-1950): forward `workspace.roots` unchanged on every leg; the `session_checkout` entry's `declared` may carry `roots` (store it as the core normalization returns it); keep the last frame's `directories` with the session row. @client-core: `StartSessionOptions.workspace` gains `roots`; project `directories`. @web-client: a per-folder isolation switch maps to `roots`; two folders of one repository share a tree ("Same worktree as …" — read `directories[].tree` after the first leg). @test (black box, wire-observable): a write to an additional folder of an isolated repository lands in the tree and its base checkout is refused; the same for a delegated child; a second repository failing at its mint refuses the leg with no tree left of the first.
|
|
576
|
+
|
|
577
|
+
### The session-deletion cascade's roster leg releases the binding of every agent spawned under the session, not only a fork child's — a plain background agent's roster entry names its tree's root and no session of its own, so the leg that selected on `sessionId` alone left it behind (#4018; behavior change: the roster leg's `selects` sentence and one more case in the deletion kit; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none — no type, key, member or export changes) (@server @test)
|
|
578
|
+
|
|
579
|
+
- **Rule.** One reading of "this session's rows" across the legs that key on a session tree: an entry belongs to a deleted session when it names that session as its session or as its root session — the two arms `BackgroundAgentStore.listBySession` already reads (the parent / owner anchor and the root anchor). The roster leg (`RETENTION_CASCADE.rosterStore`, step 6, `releaseAgent`) now selects by both.
|
|
580
|
+
- **Before.** The leg's `selects` read "every agent whose roster entry names the session as its session". Core writes `RosterEntry.sessionId` on the fork lane only (the fork child's own session); a plain background agent's entry (the Agent tool's background spawn, `src/agents/subagent.ts`) carries `rootSessionId` — the host session of its tree — and no `sessionId`. A coordinator that followed the sentence deleted a root session, deleted its background agents' rows (step 5 selects by the root anchor) and left their roster bindings: a name that still resolved, on the durable roster, to an agent whose rows were gone.
|
|
581
|
+
- **Now.** `selects` reads "…names the session as its session or as its root session (the host session of the tree the agent was spawned under — the only session a plain background agent's entry names)", and the fork row's `why` says an entry also names its root. The deletion kit's scene has one more roster agent — one whose entry names the session only as its root — and another session's such agent beside it; the kit has one more case (the roster leg releases the agent naming the session as its session and the agent naming it as its root session, with the unreadable head's shape included) and its exemption case also refuses a destructive call that names the other session's root-named agent. A fork child's entry (it names both) is released by its own session's deletion as before, and now also by its root's. `RosterEntry.rootSessionId`'s comment says the cascade reads it.
|
|
582
|
+
- **Narrowing (open to veto @server).** `sessionDeletionCascadeContract` fails a coordinator whose roster leg reads only `sessionId` (a background agent's binding stays), and one that reads every entry carrying a root (another session's binding goes).
|
|
583
|
+
- **Not changed.** `RosterEntry`'s members (none added — `sessionId` keeps meaning the retained child session, `rootSessionId` the tree's root), `RosterStore`'s members and `releaseAgent`; the leg's step, domain and verb; every other leg; what a roster row's access axes decide; the fork lane's and the background lane's writes.
|
|
584
|
+
- **Known limits.** An entry names the tree's root (and, on the fork lane, the child's own session), never the nearest spawner: (1) deleting a nested spawner session (a child, not the root) releases neither lane's bindings of the agents it spawned, though the background-agent leg's parent arm does reach their rows; (2) deleting a plain background child's own session releases nothing, because its entry carries no `sessionId`; (3) a row persisted before the root anchor existed and revived since records its parent as its root. The roster has no parent member and none is added here. The selection is not scope-filtered (the roster's `list()` spans scopes, as the one-arm reading's did): a host that shares one roster between two session stores that reuse a session id sees the other store's entries released too.
|
|
585
|
+
- **Rules.** Readings of "which rows a session tree's deletion releases" 2 → 1 (the roster leg's one-arm sentence beside the background-agent leg's two-arm one → both two-arm); special cases 0; new fields 0; the kit gains one case and one scene pair of rows.
|
|
586
|
+
- **Pins.** `test/backlog4018-roster-background-rows.test.ts` — R1 the premise over a real Agent spawn: a named plain background child's roster row names its root and no session of its own, a named fork child's names both · G1 the table's roster leg names the root session and a session-only coordinator is refused by the kit in the roster case · G1b over real rows the root's plain and fork children are released and the other tree's are not · G2 a fork child's own session still releases its row, only that row · G3 another session's rows do not move (real rows, a same-name row of another tree, and the kit refusing a coordinator that reads every entry carrying a root) · K1 a coordinator that reads the two arms clears every case of the kit. Re-read: the reference coordinators of `test/backlog3001-session-deletion-record.test.ts`, `test/backlog3005-session-deletion-cascade.test.ts` and `test/backlog4000-cascade-missing-stores.test.ts` read both arms.
|
|
587
|
+
- **Consumers.** @server (S-2560 coordinator): release a roster entry whose `sessionId` or `rootSessionId` equals the deleted session; the kit now refuses the one-arm reading. @test (black box): delete a session that spawned a named background agent — the roster listing no longer names that agent.
|
|
588
|
+
|
|
589
|
+
### What deleting a session erases names every store seat and every memory governance document and object kind — the run's persisted script and the task list mounted under the session's id are deleted with it, every governance document and control-object kind of the memory engine has an exemption row, what a leg keeps on purpose is a column (`keeps`), and an unreadable session is deleted by the same legs (`unreadable`) (#4000; behavior change: two new store verbs, three file stores compact their ledger at a session deletion, the deletion kit judges three more things; narrowing: two, below, open to veto @server; widening: none; hard-breaking: two, three sentences each, below, open to veto @server; closed sets: none; exports: none added or removed; wire: `RetentionCascadeStores` +2 required members `workflowScriptStore` / `taskListStore`, `RetentionCascadeLeg.keeps?` / `.unreadable?`, `WorkflowScriptStore.deleteRun?`, `TaskListStore.drop?`, additive except the two required members) (@server @test)
|
|
590
|
+
|
|
591
|
+
- **Rule.** One table, one reading of "what is left once a session is deleted": every store seat of `RetentionCascadeStores` and each governance document and control-object kind of the memory engine's governance store has exactly one row of `RETENTION_CASCADE`, both typed complete (the governance rows against `MEMORY_GOVERNANCE_DOC_NAMES` and `MEMORY_CONTROL_OBJECT_KINDS` — a new document or kind does not compile without its row). What a leg leaves on purpose is the leg's `keeps` column, never a comment; the operator's deletion of a row whose head cannot be read is the same legs, the leg's `unreadable` verb in its place — not a second, shorter deletion.
|
|
592
|
+
- **Before.** A workflow run's persisted script (`<scope>/<runId>.js`) had no deletion verb and outlived its run once the run and journal legs ran; a task list mounted under a root session's id had no row; the contract sentence said every governance record had a row while three of the governance store's thirteen documents and object kinds did (`session-accounts`, `settlements`, `holds`, `announcements` among those without); the session-policy high-water mark was kept by design but the row did not say so; `SessionStore.deleteUnreadable` removed the row and its log and nothing else; and the three ledger-backed file stores deleted by appending a `delete` event, the deleted row's earlier events staying in the ledger until the next housekeeping compaction (up to 1000 events later).
|
|
593
|
+
- **Now.** `WorkflowScriptStore.deleteRun?(runId, scope)` (file twin: unlinks that run's script in that scope's partition, refuses a partition that resolves outside the store, never touches the saved-name tier) is the table's step 2, found through the run store's rows like the journals; the run leg moves to 3 and every later step by one, the task list's `drop` is step 13 and the session row's leg 14. `TaskListStore.drop?()` erases the whole list with its id high-water (memory and file twins; the file twin unlinks its document and leaves the mount's directory). The governance rows are keyed by the control plane's own record names: `lineage`, `challenges`, `settlements`, `session-accounts`, `announcements`, `holds`, `origin-clearances`, `pollution`, `import-receipt`, `quarantine`, `hold`, `clearance-custody`, `capture` — each exempt, each naming the content it is a fact about. `keeps` sits on the session-policy leg (each key's high-water mark) and on both mailbox legs (the file twin's emptied box file under the box's handle). `unreadable` sits on the checkpoint leg (`deleteUnreadableSessionCheckpoints`) and the session row's (`deleteUnreadable`). `FileWorkflowRunStore.deleteBySession`, `FileBackgroundAgentStore.delete` and the file checkpoint store's two session deletions compact their ledger before they answer (`LedgerCore.settleErasure`), and also when they delete nothing while events are pending, so a repeat after a failed compaction finishes it. `sessionDeletionCascadeContract` hands the two new seats (the task-list seat is the list mounted under the session's id) and judges one more case: a session whose head cannot be read is refused before any destructive call, or deleted by every leg with the `unreadable` verbs in place, the row last.
|
|
594
|
+
- **Narrowing (open to veto @server).** A file-backed workflow-run, background-agent or checkpoint store now rewrites its snapshot at every session-tree deletion — a local executor that reuses the file family deletes sessions more slowly; the background-agent store rewrites it at every `delete()` of a row (the workflow-run and checkpoint stores once per deletion call), so a sweep that deletes rows one by one rewrites it once per row — measured by the release review, about 22 ms a row against 4 ms on 7.67.1, about 24 s against 3.8 s for 1 000 rows (KL-4000-a, #4271; @server: a deployment on the file stores), and a compaction that cannot write (an unwritable staging directory, a full disk) now fails the deletion verb instead of being swallowed.
|
|
595
|
+
- **Narrowing (open to veto @server).** `sessionDeletionCascadeContract` fails a coordinator that skips the script or task-list leg, or that starts deleting a session whose head cannot be read and then stops (a half-deleted session).
|
|
596
|
+
- **Hard-breaking — the governance rows' keys (open to veto @server).** *What changed:* `RETENTION_CASCADE`'s three governance rows of 7.67.0 are keyed `pollution`, `capture` and `lineage` (they were `session-pollution`, `session-capture-optout` and `lineage.json`, the file twin's on-disk names), beside ten new rows. *Who is affected:* a reader that looks those three rows up by the old keys — S-2560 first reads the table at this release, so no shipped consumer is known. *What to do:* read the governance rows by `heldBy: "memory-control-plane"`, or by the closed-set words `MEMORY_GOVERNANCE_DOC_NAMES` / `MEMORY_CONTROL_OBJECT_KINDS`.
|
|
597
|
+
- **Hard-breaking — the store set, the steps and the kit (open to veto @server).** *What changed:* `RetentionCascadeStores` gains two required members (`workflowScriptStore`, `taskListStore`), the session-tree steps are renumbered (journal 1, script 2, run 3, background agents' boxes 4, background agents 5, roster 6, checkpoints 7, tool results 8, rules 9, file-history carry 10 and delete 11, session box 12, task list 13, session row 14), and the deletion kit expects the two new legs and the unreadable-head case. *Who is affected:* code that builds a `RetentionCascadeStores` value, a test that pins step numbers (server `test/retention-hook-order.test.ts` pins 6–11 and 1–5), and a session-deletion coordinator run through `sessionDeletionCascadeContract` (E21). *What to do:* bind `workflowScriptStore` to `RunnerDeps.workflowScriptStore` and `taskListStore` to the list mounted under the deleted session's id; call `deleteRun(runId, scope)` for each terminal run before the run leg and `drop()` before the row; on a head read refused `corrupt_session` either refuse before deleting anything or take the `unreadable` verbs; read step numbers from `RETENTION_SESSION_TREE_STEPS` instead of pinning them.
|
|
598
|
+
- **Not changed.** The executor contract (`ManagedRetentionCapability`) and its three domains; every other leg's verb and selection as this section found it (the roster leg's selection — the session or its root session — is #4018's, above); which records are exempt (every governance row stays an exemption — none is deleted with a session); the memory control plane's records outside the governance store's two closed sets (its consolidation workbook and the sidecars it writes directly) still have no row — the sentence now claims only the two sets; `FileMailboxStore.drop` (it still empties the box file rather than unlinking it — now said by `keeps`); the housekeeping compaction cadence of every other write.
|
|
599
|
+
- **Rules.** Readings of "what a session's deletion leaves" 3 → 1 (the store rows, a comment-held exception per store, and the operator door's shorter deletion → one table with two columns); special cases 0; one new erase step shared by the three ledger stores (`settleErasure`), not three; the governance coverage moves from a sentence to a type.
|
|
600
|
+
- **Pins.** `test/backlog4000-cascade-missing-stores.test.ts` — C1 the census over real file stores: after the table-driven cascade nothing under the stores' roots names the session (path or bytes) except what a leg's `keeps` names · C2 the same for a session whose head cannot be read, through the `unreadable` verbs · C3 the script store's `deleteRun` (scope, saved-name tier, repeat, linked partition) · C4 the file task list's `drop` · C5 every governance document and object kind exempt, the session's records in them byte-identical through the cascade · C6 the table-driven coordinator clears the kit · C7 a failed compaction fails the run store's deletion and the repeat compacts. Re-pinned: `test/backlog3005-session-deletion-cascade.test.ts` (two seats, the reference reads the head, three more defects), `test/backlog3001-session-deletion-record.test.ts` (two legs, two seeded stores), `test/backlog2878-local-retention.test.ts` (seats, the script before the runs, `unreadable` verbs carried), `test/backlog514-session-incarnation.test.ts` (governance keys = the closed sets, each resolving to its file coordinate), `test/backlog1654-store-exports-retired.test.ts` (steps 1–14), the task-list kit's `drop` case, `test/backlog1468-p2-faces.test.ts` p2-delete-restart (the ledger and the snapshot name no deleted row once the call answers, where it read two `delete` events — re-pinned by a stage fix).
|
|
601
|
+
- **Consumers.** @server (S-2560 executor, E21 coordinator): new legs to run — `workflowScriptStore.deleteRun` (step 2) and `taskListStore.drop` (step 13; server's own task-list lane already drops per session); the `unreadable` verbs for an operator's deletion; the renamed governance keys; the step numbers. @test (black box): delete a session that ran a workflow — its `<scope>/<runId>.js` script is gone; delete a session with a mounted file task list — `task-list.json` is gone and the next task id is 1.
|
|
602
|
+
|
|
603
|
+
### The nested-memory lane's frame joins the witness set of a Read of the same contents — one joining rule for every delivery that records a read entry (#4038)
|
|
604
|
+
|
|
605
|
+
- **Rule.** `witnessesAfterDelivery(prior, entry, transmission)` (`src/tools/fs/fs-shared.ts`) is the one rule by which a delivery's transmission joins an entry's witness set: a delivery of the same contents in the same window as the prior entry joins its set, any other delivery starts the set over. A whole delivery (`truncated: false`) is one window whatever line numbers its entry records — a Read counts the ghost line a trailing newline renders, an instruction-file frame does not.
|
|
606
|
+
- **Before.** The Read tool's commit joined the set; the lane's record wrote its own `witnesses: [frame]` over whatever entry stood. A frame that landed after a Read of the same contents turned the entry into one that only the lane's frames witness (no longer an access for the lane's law), and the frame's later trim retired an entry a Read result still carried — the edit gate then asked for a fresh read.
|
|
607
|
+
- **Now.** The lane's record takes the same rule: the frame joins the Read's witnesses, the frame trimmed leaves the entry standing (the Read's result still carries the body), and the Read's departure retires it. While the frame is the newest witness the entry reads as the lane's (`isInstructionLaneEntry` reads the newest witness), so it is not an access for the lane's law until a later Read joins.
|
|
608
|
+
- **Widening (open to veto).** The Read tool's own join now also treats two whole deliveries as one window when their recorded line numbers differ (a transcript-seeded entry, which records no window, and a whole Read join; before, the second delivery started the set over). The join only keeps an entry standing while a transmission that carries the body is still in the request.
|
|
609
|
+
- **Not changed.** The retirement arms (a trim, a clear, the budget's pointer), compaction's newest-witness rule, a delivery of another body or another window (starts over), the wire (no new key).
|
|
610
|
+
- **Rules.** Retired: the lane's own set-building (two readings of "which transmissions carry this body" ⇒ one). Added: none.
|
|
611
|
+
- **Pins.** `test/backlog3889-nested-memory-lane.test.ts` P5b (the frame joins a Read's set; the trims retire in order) · P5c (another body starts over) · P5d (another window starts over); `test/backlog3942-witness-set.test.ts` unchanged and green.
|
|
612
|
+
- **Consumers.** None (engine-internal).
|
|
613
|
+
|
|
614
|
+
### Two readings fail closed where an answer outside the contract used to escape them — `openBoundAt` refuses a `null` member, a throw and a non-result answer with `unknown`; the restore seat reads an empty or blank code as `unknown` (#4040)
|
|
615
|
+
|
|
616
|
+
- **Rule.** A reading of a host adapter's answer has no third kind of outcome for an answer it was not written for: it is the refusal the reading already has, or the code the reading already has for "no fact named". `openBoundAt` (`src/tools/fs/bound-open.ts`) never throws; `establishAtRestoreSeat` (`src/core/runner/restore-seat.ts`) reads only a code the answer names.
|
|
617
|
+
- **Before.** `openBoundAt` called `env.openBound` bare: an adapter's `null` member (a JS adapter without types), a throw, or a value that is no result raised out of it. The plan file's reader and the attachment reader met an exception neither site catches; the LSP session's gate returned the raw thrown text as its reason, and the LSP target facts and the nested-memory lane each read it as "no answer" in their own way. The restore seat read an establishment failure whose `code` was empty as the environment's standing refusal and failed the leg as `connect refused ()`.
|
|
618
|
+
- **Now.** Those answers — a `null` member, a member that cannot be read, a throw or a rejection, a value that is no result (no `ok`, an `ok` without an object, a failure without a code) — are `{ kind: "refused" }` with the code `unknown` and the adapter's words bounded to one line by `envErrorDetail` (the readers' own refusal path; no reader reads the object). An empty or blank `code` names no fact of the environment: it reads as `unknown` and the run starts, as a connect that throws or answers outside its shape already did.
|
|
619
|
+
- **Narrowing (open to veto @server).** An adapter whose `openBound` throws or answers outside its type now reads as the one refusal at the plan file's reader, the attachment reader and the LSP gate and target facts (the object is not read): the first two no longer raise, the LSP gate's reason is the bounded line instead of the raw thrown text. The nested-memory lane keeps the by-path read it had for any answer but a bound object (a refusal there still reads the file by its path until that lane's own change). The three catches that read a throw their own ways (`src/core/lsp.ts` twice, the lane's `openBound` call) are now unreachable and stay for the changes that own those lines.
|
|
620
|
+
- **Widening (open to veto @server).** A leg whose adapter's `connect` failure carries an empty or blank `code` now starts (its first operation meets the environment) instead of failing at the restore seat; a `resumeVM` failure with such a code is answered back to the resume's own reading, `resume.env_failed` (as an `unknown` always was), instead of the seat's own refusal.
|
|
621
|
+
- **Not changed.** `not_supported` and the absent member (unbindable), every named code's reading, `RETRYABLE_REMOTE_ERROR_CODES`, the connect arm's contract check, the wire (no new key; `unknown` is an existing `FileErrorCode`).
|
|
622
|
+
- **Rules.** Retired: the two unclassified shapes (an exception from the reading; a standing refusal that names nothing). Added: none — both fold into the existing refused / `unknown` readings.
|
|
623
|
+
- **Pins.** `test/backlog4040-fail-closed-readings.test.ts` — A1 a `null` member · A2 a throw, synchronous or rejected · A3 a thrown value that is no Error · A4 the message bounded to one line · A5 a non-result answer (six shapes) · A6 controls (absent, `not_supported`, bound, another code) · A7 a member that cannot be read · B1 an empty code answers back · B2 a blank code · B3 a named code outside the table fails the leg · B4 `unknown` / `aborted` / retryable answer back · B5 `connectAtRestoreSeat` · R1 the whole run starts · R2 a standing code still fails in prepare · R3 a resume whose `resumeVM` fails with an empty code ends `resume.env_failed`.
|
|
624
|
+
- **Consumers.** @server (JS adapters): an `openBound` or establishment call that breaks its contract is answered with the refusal / `unknown` reading instead of an exception / a standing refusal.
|
|
625
|
+
|
|
626
|
+
### The brain layer's `toolcall_end` comments say what the event is now — a stream event the loop forwards to its consumers, with tools starting once the stream has ended — and no longer name the retired in-stream executor (#4042)
|
|
627
|
+
|
|
628
|
+
- **Rule.** `toolcall_end` says a call's input is complete: emitted at most once per call, only when the input parsed to a valid object; every emitted call appears in the `done` message (an `error` final may lack it), and a call closed at finalize emits none. The loop forwards it to its consumers as a `message_update`; the tool starts after the stream has ended.
|
|
629
|
+
- **Before.** The Anthropic, OpenAI and Open Responses parsers, the stream engine's retry tiering and the provider contract on `AssistantMessageEvent` still gave an in-stream executor as the reason a call closes early, a malformed close "never executes in-stream", and a re-send after a completed call "executes it twice".
|
|
630
|
+
- **Now.** Each of those sentences says what the code does today (a malformed close tells no consumer a complete call exists; a re-send would replace a response whose completed call consumers have already seen; the early close of lower slots streams the call while the response goes on).
|
|
631
|
+
- **Not changed.** Every line of behaviour and the wire — comments only (`src/brain/anthropic.ts`, `openai.ts`, `open-responses.ts`, `stream-engine.ts`, `src/engine/llm/types.ts`).
|
|
632
|
+
- **Rules.** None. **Pins.** None (comments; the comment-ratio ratchet's measured rows). **Consumers.** None.
|
|
633
|
+
|
|
634
|
+
### The rewind agreement check reads the key a read record was judged under; three metadata-only reads are a known limit — `KL-4048-a` (#4048)
|
|
635
|
+
|
|
636
|
+
- **Rule.** The check that decides whether a restored file's read record still agrees with the restored bytes reads the file at the key the record is held under (`k`, the key the noted path was judged to), never at the spelling the restore noted.
|
|
637
|
+
- **Before.** The check judged the noted path to a key, looked the record up by the key, then read the file by the noted spelling — two names for one decision.
|
|
638
|
+
- **Now.** One name: the judged key. A restored path that became a link to a file the session read is compared against that file, the file the record stands for.
|
|
639
|
+
- **Not changed.** What the check keeps and drops, the restore channel, the wire. `KL-4048-a` records that the changed-file hint (mtime), the read station's card kind and the write tool's unread-refusal size bit read metadata by key after their own judgment and carry no file bytes.
|
|
640
|
+
- **Rules.** Retired: the spelling read. Added: none.
|
|
641
|
+
- **Pins.** `test/backlog1130-rewind-files-verb.test.ts` — "the agreement check reads each restored file at the key its record is held under …" (a restored path that became a link to a read file; the read is at the key).
|
|
642
|
+
- **Consumers.** None (engine-internal).
|
|
643
|
+
|
|
644
|
+
### The Edit tool's missing-file refusal names no sibling under a read-deny row — its "Did you mean" names go through the read face's table as Read's do (#4049)
|
|
645
|
+
|
|
646
|
+
- **Rule.** A "Did you mean …" name is an existence disclosure; it is judged by the read face's deny table wherever the engine offers one (`enoentMessage`'s `readDeny` seat).
|
|
647
|
+
- **Before.** Read's missing-file refusal filtered every candidate name (and refused to list a deny-listed directory); Edit's passed no table, so a missing name next to a deny-listed file (`id_rsa.txt` beside `id_rsa.pub`, a twin of `.netrc` in the cwd) named it.
|
|
648
|
+
- **Now.** `createEditFileTool` takes the read face's table as a last optional seat (the toolkit passes it) and hands it to the refusal: the same candidates are filtered the same way.
|
|
649
|
+
- **Narrowing (open to veto @test).** An Edit refusal for a missing name omits the suggestion when the only candidate lies under a read-deny row.
|
|
650
|
+
- **Not changed.** The refusal's first sentence and cwd note, Write and NotebookEdit (they create or refuse by their own arms), the wire.
|
|
651
|
+
- **Rules.** Retired: Edit's unfiltered suggestion. Added: none (the one existing filter, one more caller).
|
|
652
|
+
- **Pins.** `test/read-deny-faces.test.ts` — E1 same-stem names (both sides of the dot) · E2 the cwd-rebase twin · E3 an ordinary sibling still suggests · E4 Read's refusal unchanged.
|
|
653
|
+
- **Consumers.** @test (the sibling name no longer rides Edit's refusal under a deny row).
|
|
654
|
+
|
|
655
|
+
### An `openBound` answer of `not_supported` reads as the method's absence on that target — every reader that binds a read takes its one no-binding disposition for that code and for no other (#4035; ruled C-R1357; behavior change: the plan file's reader reads by name where the environment answers `not_supported` (it refused), and a Linux host that cannot read `/proc` answers `not_supported` from `openBound` (it answered `precondition_failed`); narrowing: none; widening: one, below, open to veto @server; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli)
|
|
656
|
+
|
|
657
|
+
- **Rule.** An environment's `openBound` answering `not_supported` on a target is that method's absence on the target — one fact. One judgment (`openBoundAt`) reads every answer as `bound`, `unbindable` or `refused`: `unbindable` is the member missing or answering `not_supported`, and only that falls back to a reader's no-binding disposition; `precondition_failed` and every other code are refusals that never fall back. The readers are the file reader behind a user attachment (`cannot_bind`), the session plan file's reader (its by-name arm), the LSP session's read gate and the LSP tool's target facts (as #4035 wrote them, one refusal sentence and nothing to say; #4034 S1a in this release, below, makes both read through the one simulation — `readObjectAt` — so an environment that cannot bind has its LSP reads and target facts answered by the simulation); no other source reads `env.openBound`.
|
|
658
|
+
- **Before.** Four readers, three readings. The attachment and the LSP target facts read `not_supported` as absence; the plan file's reader refused a read on it (only a missing member took its by-name arm); the LSP gate refused in a sentence of its own for a missing member and in the environment's for `not_supported`. The node environment's Linux arm answered `precondition_failed` — "The path named another object than the one opened" — when it could not read the kernel's path of the opened descriptor (no `/proc`), a reason that was not true.
|
|
659
|
+
- **Now.** The node environment's Linux arm answers `not_supported` when the descriptor's kernel path cannot be read (the opened object cannot be placed) and `precondition_failed` only when the path it reads differs. The contract's sentences say the reading — `BoundFileObject` and `FileSystem.openBound`, the attachment seat in the integration guide, the LSP read admission and its target facts, the attachment reader, the optional-method census: every answer is about the bound object or is a refusal; no answer comes from an object reached through a link; the object is placed at the path by one observation; an environment that binds again on every answer refuses after a rename or a removal and may wait on a FIFO until its own timeout; equal identities across two opens do not prove one object once an inode is reused; `not_supported` is read exactly as the member's absence.
|
|
660
|
+
- **Widening (open to veto @server).** The plan file's reader now takes its by-name arm (judged, read, judged again) on a target whose environment answers `not_supported` — including a Linux host that cannot read `/proc` — where it refused the read. It is the arm an environment with no `openBound` already takes.
|
|
661
|
+
- **Reading (@server ssh leg, @cli).** A remote environment that cannot bind a target (a non-Linux target) answers `not_supported`: an attachment is `cannot_bind` exactly as for a missing member, the plan file is read by name, and an LSP read takes the same one by-name simulation (#4034 S1a, the same version). A refusal that is not `not_supported` (`precondition_failed`, `invalid`, …) keeps its own reason and never falls back.
|
|
662
|
+
- **Not changed.** The wire (`not_supported` is already a `FileErrorCode`; no new key); a Linux host that can read `/proc` (every answer as before); the by-name arm's own judgments; the reasons of every other code.
|
|
663
|
+
- **Rules.** Retired: three readings of one code, the file reader's own synthesized `not_supported`, a Linux answer whose reason was untrue. Added: none — one judgment the four readers call, in place of four readings.
|
|
664
|
+
- **Pins.** `test/backlog4035-not-supported-one-reading.test.ts` — G1 the plan reader reads by name on `not_supported`, as with a missing member · G2 every other code refuses with its reason, no by-name read · G3 the LSP gate takes one disposition for both forms (since #4034 S1a the by-name simulation; it refused in one sentence), the target facts read both alike · G4 the attachment is `cannot_bind` for both forms, another refusal is its own card · G5 Linux: unreadable `/proc` ⇒ `not_supported`, a differing path ⇒ `precondition_failed` · G6 the nine contract sentences (and the two docs lines that still said "held open"), and no source but `bound-open.ts` reads `env.openBound` or the code.
|
|
665
|
+
- **Consumers.** @server (S-3356 ssh leg): answer `not_supported` on a target the leg cannot bind; the engine reads it as a missing member. @cli: an attachment an environment cannot bind states `cannot_bind` whichever way the environment says so.
|
|
666
|
+
|
|
667
|
+
### A leg's own remote environment is established once, at its restore seat, before anything reads it — `connect` (or `resumeVM` on a leg that restores a snapshot) with the run's cancel signal — and an answer the retryable table does not hold fails the run in prepare with that code (#4039, #4025 S1; ruled C-R1356 + addenda 1 / 2; behavior change: the engine calls `RemoteExecutionEnv.connect` on every leg that restores no snapshot, a standing refusal ends the run before its first turn, a cancel at the seat ends it there; narrowing: two, below, open to veto @server @cli @client-core @test; widening: one, below, open to veto; hard-breaking: two, three sentences each below, open to veto @server; closed sets: none — the union `RemoteExecutionErrorCode` and `RETRYABLE_REMOTE_ERROR_CODES` keep their members, the field beside them widens; exports: none added, none removed; wire: two member types widen (`RemoteExecutionError.code`, `RemoteEnvFailureNote.code`) and two signatures widen a parameter (the `RemoteExecutionError` constructor, `isRetryableRemoteErrorCode`) — the four BREAKING rows are hard-breaking ②; no closed set moved, none unreadable) (@server @cli @client-core @sdk @test)
|
|
668
|
+
|
|
669
|
+
- **Rule.** The remote environment the leg's factory minted for this leg (past the worktree guard; `isRemoteExecutionEnv`) is established once at the leg's restore seat (`prepareWorkspaceRestore`, before the settle point reads the leg's target, the session checkout reads its tree and the task root is canonicalized): `resumeVM` on a leg that restores a snapshot (its bounded retry unchanged), `connect({ abortSignal })` on every other leg — a fresh leg, a delegated child, a resume without a workspace handle, a park-only resume after its identity check. One reading of the answer, three steps: ① the run's own signal is aborted — read before the call and again once it settled — ⇒ the leg stops with the run's one cancellation terminal `{ kind: "failed", message: "run aborted" }`, no code; ② `ok` ⇒ the run starts; ③ a failure is read through `RETRYABLE_REMOTE_ERROR_CODES` (its third reader, beside the retry decision and the failure note's `retryable`): a code it holds, `unknown`, `aborted` the run's own cancel did not cause, and a `connect` that throws or answers outside its contract (read as `unknown`) answer back to the verb's own reading (`connect`: the run starts and its first operation meets the environment as it would have; `resumeVM`: `resume.env_failed` with its note, as before); **any other code is this environment's standing refusal** — `auth_failed`, `unsupported`, `outcome_unknown`, `suspended`, `command_in_flight`, `post_resume_failed`, a deployment's `env.*` — and the leg fails in prepare with that code as `terminal.code`, `terminal.message` = `<verb> refused (<code>): <the adapter's sentence>`, zero turns, zero tokens, no `remoteEnvFailures` note, the environment destroyed by the prepare rollback. Starting a run past a failed establishment is itself a retry (its first operation establishes again), so the seat splits where the retry table splits.
|
|
670
|
+
- **Before.** The engine called `connect` nowhere: a fresh leg's first fs / exec touch connected lazily, and a refusal met there was read as "the root did not canonicalize" and the run started. One fact — the environment refuses this run — had three terminals: a fresh leg `completed` with the refusal only in tool results, a session with a checkout record `workspace.checkout_unavailable`, a snapshot resume `resume.env_failed`. A cancel while `resumeVM` was in flight ended `resume.env_failed`; one at the prepare of a fresh leg on a remote environment let the checkout and the canonicalization reach the environment first and ended `run aborted` at the leg's entry.
|
|
671
|
+
- **Now.** One terminal per fact: the standing refusal's own code on every leg; the run's cancel `run aborted` (no code) wherever the run notices it (one mint, `runAbortedError`, shared by the seat and the leg's entry).
|
|
672
|
+
- **Hard-breaking — two, three sentences each (open to veto @server).** ① *What changed:* the engine calls `RemoteExecutionEnv.connect` once on every leg that restores no snapshot, at the restore seat, with the run's cancel signal; its contract now says it MUST be idempotent — on an environment already connected it answers the current handle and provisions nothing — and a connect that answers a failure leaves nothing a second connect would duplicate. *Who is affected:* every adapter, because the seat is commonly not the environment's first connection (a delegated child's placement check touches it first, a factory may write into it before handing it over, a parked leg's factory may hand back a live environment); the server's k8s `connect` joins only a connect in flight (`remote-env-k8s.ts:1029`), so a second connect on a connected environment provisions a second pod and leaks the first. *What to do:* give k8s `connect` an already-connected guard in the same pickup as this release, and check the other lanes (e2b, ssh, local-docker and host already answer the current handle; device single-flights; adb re-handshakes without leaking). ② *What changed:* `RemoteExecutionError.code` — its `constructor` parameter, `RemoteEnvFailureNote.code` and `isRetryableRemoteErrorCode`'s parameter with it — is `RemoteExecutionErrorCode | \`env.${string}\``, so an adapter spells a refusal the closed set has no word for (a bucket that is gone, an occupant that is missing) in the `env.` namespace and it reaches the terminal verbatim. *Who is affected:* a TypeScript consumer that uses one of those member types as the KEY of an exhaustive record or switch (none found in the server; its three device tables use it as a value type and stay green). *What to do:* add an `env.` arm where you switch on the code; an `env.` code is never retryable, folds to `config` in `errorClassOf` and to `environment` in `resumeStartupCauseOf`.
|
|
673
|
+
- **Narrowing ① (open to veto @server @cli @client-core @test).** A run whose environment answers a standing code at the seat no longer starts; it fails in prepare with that code (before: `completed`, `workspace.checkout_unavailable` or `resume.env_failed`). Four server lanes change without a server edit, because they already spell such a refusal with a closed-set permanent word: device lifecycle refusals (`auth_failed` — revoked, claim conflict, identity mismatch, …), an ssh key the host rejects (`auth_failed`; one failed authentication per run, where it was one plus one per tool), an adb device without `shell_v2` (`unsupported`) and adb's catch-all `auth_failed`, k8s pod creation answered 401 / 403 (`auth_failed`, the quota 403 among them). `workspace.checkout_unavailable` no longer carries an environment's refusal (the seat precedes the checkout); `resume.env_failed` no longer carries a `resumeVM` standing refusal.
|
|
674
|
+
- **Narrowing ② (open to veto @server @cli @test).** A resume whose `resumeVM` answers a closed-set standing word: the parked-resume receipt's `cause` is absent where it read `environment` through `resume.env_failed` (KL-4039-a); a resume its own run cancels while `resumeVM` is in flight ends `run aborted` (no code) where it ended `resume.env_failed`. The reopen reason is `env_failed` in both, as before.
|
|
675
|
+
- **Widening (open to veto).** The seat precedes the settle point, so an adapter whose `workspaceHandle()` throws until it is connected now reports its target there: what its leg spawns carries a fence, and a leg fenced to that device runs where it was refused `config.placement_not_honored`. The server's lanes answer the handle before connecting (no change there); a third-party lazy adapter is affected. A resume leg is not widened: its identity check reads the handle before it establishes, as before (a workspace that is not the checkpointed one is never connected to).
|
|
676
|
+
- **Deviation from the reference (D-4025-a, open to veto).** The reference refuses to start when its startup sandbox initialization throws (`cli316.js` L288946–L288948; an initialization repeated after a settings change that fails is only logged, L288950); the seat starts the run past a retryable or unclassified establishment failure, because a remote lane establishes again at the run's first operation. A standing code refuses to start, as the reference does.
|
|
677
|
+
- **Not changed.** The members of `RemoteExecutionErrorCode`, `RETRYABLE_REMOTE_ERROR_CODES` and `ResumeStartupCause`; `resumeVM`'s bounded retry; a `resumeVM` that throws passes through prepare as before; `postResumeInit` failures stay `resume.env_failed` with their note; a deployment's static env, a local env and the stub are never established; the worktree guard runs before the seat; the session checkout and the task-root canonicalization still read the environment without a signal (a cancel that lands while they wait is answered at the leg's entry, as before).
|
|
678
|
+
- **Rules.** Added: one seat (one reading, `establishAtRestoreSeat` in the machinery module `restore-seat.ts`; the verb by leg), one more reader of the existing retryable table (its two exclusions are the table's own "says nothing of the environment" classes, `unknown` and caller-driven `aborted`), one cause-table arm (`env.` ⇒ `environment`), one field widening with no new name. Retired: the second copy of the cancellation mint (one mint, `runAbortedError`, `terminal-cause.ts`); two terminals for one cancel; three terminals for one standing refusal; "the engine never calls `connect`". No list of lanes or codes was added.
|
|
679
|
+
- **Pins.** `test/backlog4039-prepare-env-seat.test.ts` — P1 a fresh leg's standing refusal, seven codes, ends in prepare with the code, the adapter's sentence, zero model calls and the env destroyed once · P2 the nine answers the run starts past (five retryable, `unknown`, an adapter's own `aborted`, a throw, a value outside the contract), connect exactly once · P3 a cancel during connect reaches its signal and ends the leg `run aborted` with nothing after connect · P4 connect is a top-level fresh leg's first operation · P5 a snapshot resume's `env.*` / `auth_failed` is the terminal code, no note, one attempt, the row pending again under `env_failed` · P6 a session with a checkout record answers the refusal, not `workspace.checkout_unavailable`, and its tree is never read · P7 a static remote env and a non-remote factory env are never connected · P8 park-only: connect once; its standing refusal is the terminal; a wrong backend is refused before any connect · P9 the cause table reads `env.*` as `environment`, `auth_failed` as absent · P10 `env.*` is never retryable · P11 the worktree guard detaches an aliased static env before the seat · P12 a delegated child's env touched by its placement check is still connected once (`connectWhileConnected === 1`) · P13 a handle that answers only once connected honors its device fence · P14 the cancelled run's terminal is exactly `{ failed, "run aborted" }` on a fresh leg (connect never called, nothing reached the env) and on a snapshot resume cancelled in `resumeVM` · P15 a `resumeVM` that throws passes through. `test/rb439-remote-env-lifecycle.test.ts` re-pinned: `auth_failed` on `resumeVM` is now the terminal code with no note; `unknown` stays `resume.env_failed` with its note. The fixture (`test/mock-remote-env.ts`) gains five optional cells, defaults unchanged: a widened outcome code, `connect` outcome / `"throw"`, `onConnect`, the `ops` order, the connected state with `connectWhileConnected`.
|
|
680
|
+
- **Consumers.** @server (S-3449, the same pickup): the k8s `connect` already-connected guard (hard-breaking ①); spell each remembered refusal with a non-retryable code — k8s `holdRefusal` / `refusePod` / the unreadable-occupant refusal and e2b's connect refusal, adb's home refusal and ssh's host-key mismatch all answer `connect_failed` today, so the run keeps starting and meets them again at its first operation — `env.snapshot_bucket_missing` / `env.occupant_missing` / `env.occupant_unreadable` or a closed-set permanent word; `neverOpened` (`remote-env-file-error.ts`) compares `=== "connect_failed"`: read the new `env.*` codes too, or a refusal met after the run started falls to `unknown` and the dead-directory diagnosis covers the refusal's sentence; check the spelling of device `claim_conflict` on the lifecycle face (`auth_failed`, retryable on the exec / file faces), the k8s quota 403, adb's catch-all `auth_failed` and ssh's internal `aborted`; the k8s comment that core never calls `connect()`. @sdk: the "`env.*` has one member today" sentence. @cli: a failed run's code may now be `auth_failed` / `unsupported` / … or an `env.*` code (`errorClassOf`: `env.*` ⇒ `config`, the closed-set words ⇒ `unknown`). @client-core: the same terminal codes reach the run's terminal frame. @test: HB1–HB8 (the black-box criteria handed with this car).
|
|
681
|
+
|
|
682
|
+
### The loop has no in-stream tool lane — `StreamToolExecutor`, `AgentLoopConfig.streamingToolExecution` and the late-cancellation latch only it read are retired, so a tool call starts after its stream's final at the loop as at the Runner (#4032; behavior change: none on any Runner leg; a harness built with `streamingToolExecution: true` and no `tool_call` handler ran the lane in 7.67.1 (the harness forwarded the switch), and that option is gone from `AgentHarnessOptions` with #3830 in this release, so the lane's last reader was a caller handing the raw loop the switch itself; narrowing: a raw `runAgentLoop` caller still writing the switch now gets the batch order; widening: none; hard-breaking: none; closed sets: none; exports: none — `runAgentLoop` is not on the export map; wire: none) (@test)
|
|
683
|
+
|
|
684
|
+
- **Rule.** A tool call starts after the stream's final: preflight, execution and results all follow it. The loop kept an executor that started a concurrency-safe call at `toolcall_end` while the model was still streaming, behind `AgentLoopConfig.streamingToolExecution`. The Runner's member and the harness option that forwarded it are retired (#3830), so nothing in the package set the switch any more: the lane was reachable only by a caller handing the raw loop the switch itself.
|
|
685
|
+
- **Before.** With the switch on, the loop carried a second set of rules beside the batch pipeline: an admission predicate with a source-order barrier and the concurrency cap, a settle barrier (a `try`/`finally` at each of the three stream sites), a harvest of the held results at the final (batch pipeline, errored final, interrupted turn, orphan close), a merge-back of the called blocks into an errored final, a latch forwarding a cancellation that rejected a progress delivery after the call's outcome was decided, and guards in the prompt-too-long / refusal / degenerate / blank-failure arms that read the executor's admitted count.
|
|
686
|
+
- **Now.** The class, its three helpers, the config member, the raw executor's late-cancellation parameter, the three settle barriers, the harvest in the interrupted-turn settlement and the admitted-count conjuncts are gone (`agent-loop.ts` 3565 lines in 7.67.1 → 3136 in 7.68.0, other sections' edits of the file included). `executeToolCallsPartitioned` is the partition loop and nothing before it. A config that still carries the key reads nothing: the call starts after the stream's final.
|
|
687
|
+
- **Not changed.** The order a Runner leg runs tool calls (it never used the lane); the batch pipeline's partitioning, preflight, concurrency cap and halt directive; the interrupted-turn settlement's paired results for never-started calls; the update-delivery band (a rejected delivery never rewrites a committed result; a wrapped cancellation stays a hard failure on the batch path).
|
|
688
|
+
- **Consumers.** @test: sema-test's direct `runAgentLoop` repros that set the loop's `streamingToolExecution` now read the batch order (the call starts after the stream's final); the loop is a dist-internal path, not a package face, so this is not a hard break. The brain layer's early `toolcall_end` emission and the comments that name the retired executor (`src/brain/anthropic.ts`, `openai.ts`, `open-responses.ts`, `stream-engine.ts`, `src/engine/llm/types.ts`) are untouched here — the provider contract's rationale is a separate question.
|
|
689
|
+
- **Rules.** Retired (−7, added 0): the opt-in switch · the admission predicate with its barrier and cap · the settle barrier · the held-result harvest (four seats) · the errored-final merge-back · the late-cancellation latch · the admitted-count guards on the recovery arms.
|
|
690
|
+
- **Pins.** `test/backlog4032-stream-tool-executor-retired.test.ts` — S1 no file under `src/` names the retired executor, its helpers, the switch or the latch · B1 a safe call whose arguments completed mid-stream starts only after the final, also for a config still carrying the key · B2 an errored final carrying such a call: the call never ran, the run ends on the failure · C1 control, no key, same order · B3 control, a batch's results are committed before the next batch's gate runs (the order the lane's own harvest held, kept by the pipeline that remains). Baseline (the branch's merge with the retirement of the task member, `de21a6b64`) red on S1 / B1 / B2 (B1: the call started before the final, order `tool:start`, `final`), controls C1 / B3 green; green here. Re-keyed to the batch path (the law each held survives the lane): `test/backlog2684-streak-stream-interrupt.test.ts` (the malformed streak is cleared by a turn the user's interrupt ends on its aborted final; the call never started), `test/human-reject-batch-boundary.test.ts` (the first batch's block + halt keeps the later sequential call from starting), `test/tool-update-delivery-disclosure.test.ts` (the raw executor has one caller). Retired with the lane: `test/streaming-tool-execution.test.ts` (thirteen cells), the three in-stream cancellation cells of `test/tool-update-delivery-disclosure.test.ts`, the in-stream settle-barrier cell L1 of `test/design384-slice3-halt-attribution.test.ts` (its law — a cut landing after the final was obtained does not claim it — stays pinned by L1b), the loop control C1 of `test/backlog3830-streaming-tool-execution-retired.test.ts`. Mutations, each red on its pin: M1 a retired name in a src comment (S1) · M2 a safe call started at `toolcall_end` (B1, B2, and C1 — the mutation ignores the switch) · M3 the malformed streak not cleared on the aborted final's exit (`backlog2684`) · M4 the first batch's block + halt not raised as the pipeline's boundary (eleven cells of `human-reject-batch-boundary`, the re-keyed one among them) · M5 a second caller of the raw executor (`tool-update-delivery-disclosure`) · M6 a batch's results not delivered where the next gate can count them (B3).
|
|
691
|
+
|
|
692
|
+
### An approved edit of the question tool, re-judged into a question, takes the content route — the engine's question call meets the same channel at the re-judge that it meets at the first adjudication (#4037; behavior change: yes, host-only — a question call whose edit an inherited installation's person approved; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core)
|
|
693
|
+
|
|
694
|
+
- **Rule.** A question the engine's own question tool puts to a person is a content ask: a permission approver's yes / no cannot carry an answer (an allow would run the tool into a face nobody answered, a deny would tell the model a question was shown that never was). It is delivered through the run's content-ask channel, one delivery and one reading of its outcomes wherever the question is raised: answered ⇒ the call runs on the bytes the question presented; delivery failed after the question may have been shown ⇒ it runs too (the failure surfaces through the tool); nobody to show it to ⇒ the refusal that says no live face answered it, and a park only where the park stepped aside for the face.
|
|
695
|
+
- **Before.** The first adjudication sent the question there; the re-judge that follows a person's approved edit of the call's arguments (an inherited installation's person edited the call) put the edit's new question to the permission chain whatever the call was.
|
|
696
|
+
- **Now.** In the re-judge, a question on the engine's question tool (judged off the call's face, never its name) goes to the content channel: `putContentAsk` and `contentAskNotShown` are the one delivery and the one refusal both seats read, with the settlement the channel's word (`approver_unavailable` when a channel was wired and said so, `no_approver` when none is). A plain tool's re-judge is unchanged: it asks the permission chain.
|
|
697
|
+
- **Behavior change (open to veto @server @cli @client-core).** On a durable question policy, an installation approving an edit of a question call used to put the edited question to the permission approver; it now reaches the question face, or refuses saying no live face answered it. The entrance is host-only (an inherited installation approving an edit of the question call); no Runner configuration of the package's own produces it.
|
|
698
|
+
- **Not changed.** The first adjudication (sentence, settlement, order of ledger and deny attribution byte for byte); the re-judge's own mode read point (it keeps its inline arbitration beside the first adjudication's; folding the two is a separate decision); every plain tool's approved-edit round.
|
|
699
|
+
- **Rules.** The content route's delivery and its not-shown refusal were each spelled once in the first adjudication; they are two closures now read at two seats (−2 inline copies, +0 rules), and the re-judge gains an arm that calls them (no new sentence, no new settlement).
|
|
700
|
+
- **Pins.** `test/backlog4037-edit-round-content-ask.test.ts` — E1 answered ⇒ allowed on the edited bytes, the channel handed the edited request once, the permission chain never asked · E2 unavailable ⇒ the no-live-face refusal, settlement `approver_unavailable`, the first adjudication's sentence and record · E3 no channel ⇒ same sentence, `no_approver` · E4 unavailable with the park stepped aside ⇒ one flagged park attempt, the suspension is the result · E4b unavailable with the park not stepped aside ⇒ no park attempt, the refusal · E5 delivery failed ⇒ allowed on the edited bytes, never re-routed · K1 control, a plain tool's re-judge asks the permission chain once on the edited bytes · K2 control, a caller's tool named `AskUserQuestion` (no engine face) asks the permission chain. Baseline (`de21a6b64`) red on E1 / E2 / E3 / E4 / E4b / E5, controls K1 / K2 green; green here. Mutations, each red on its pin: M1 the re-judge never takes the content route (E1–E5, E4b) · M2 judged by the tool's name, not the engine's face (K2) · M3 the park re-entered whether or not it stepped aside (E4b) · M4 a failed delivery read as nobody-to-show-it-to (E5) · M5 the re-judge's refusal in its own words (E2, E3, E4b).
|
|
701
|
+
|
|
702
|
+
### The MCP authorization sentences read one source for where the user authorizes — #4029's neutral helper is superseded by #3984's fixed phrase `WHERE_TO_AUTHORIZE`, read by the frame's needs-authorization arm, the call-time refusal line and the host notice alike (#4029; behavior change: none of its own — the sentences change with #3984 in this release, above, and #4029 adds nothing to them; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@test)
|
|
703
|
+
|
|
704
|
+
- **Rule.** Where the user authorizes is one fact read from one source, not a sentence of any frame. Three places say it: the `mcp_unavailable` frame's needs-authorization arm, the call-time refusal line (`unauthorizedLine`, through `userMustAuthorize`) and the host notice (`mcpAuthorizationNotice`).
|
|
705
|
+
- **Before.** The branch renamed the shared helper `whereToAuthorize(url)` to a neutral `authorizationAddressOf(url)`, still spelling the seat's URL.
|
|
706
|
+
- **Now.** #3984 (this release, above) retires the helper (7.67.1 still has `whereToAuthorize`): no authorization sentence spells an address, and the three places read one fixed phrase, `WHERE_TO_AUTHORIZE` (`src/core/mcp-failure.ts`). The rename does not land; #4029's ask — one source no frame owns, no catalog id in it — is what that phrase already is. Zero behavior.
|
|
707
|
+
- **Rules.** Retired: none here (the frame-flavoured name leaves with #3984). Added: none.
|
|
708
|
+
- **Pins.** `test/backlog4029-authorization-address-neutral.test.ts`, re-read onto `WHERE_TO_AUTHORIZE` — A1 the old name is gone from `src/` · A2 the phrase is referenced from exactly the arm, `userMustAuthorize` and the host notice · A3 its declaration carries no catalog id · A4 the frame's own wording lives in the arm alone.
|
|
709
|
+
|
|
710
|
+
### `TaskSpec.streamingToolExecution` is retired — a Runner leg never started a tool call before its stream final, so the opt-in that promised it goes, with the harness option it was forwarded through; a spec still writing it is refused `config.spec_unknown_key` (#3830; behavior change: none for a run that does not write the key; narrowing: a spec that writes it is refused instead of run, counted as the hard break below; widening: none; hard-breaking: one, three sentences below, open to veto @server @cli @client-core; closed sets: none; exports: none; wire: `TaskSpec.streamingToolExecution?` removed) (@server @cli @client-core)
|
|
711
|
+
|
|
712
|
+
- **Rule.** A Runner leg runs every tool call after its stream final, and no seat says otherwise. The loop's own in-stream lane (`AgentLoopConfig.streamingToolExecution`, internal, not exported) had no writer once the seats above were retired, and is retired in the same release (#4032, below).
|
|
713
|
+
- **Before.** Every Runner leg folds the always-on integrity guards and the permission-mode gate into its tool policy, so every leg registers the `tool_call` gate, and the harness forwarded the opt-in to the loop only when no `tool_call` handler was registered. The option never took effect on any Runner run, while its docstring promised in-stream execution and said a gated task would simply ignore it — every task is gated.
|
|
714
|
+
- **Now.** `TaskSpec` has no such member and `AgentHarnessOptions` no such option. A spec still writing the key (`true` or `false`) is refused at prepare by the member door (`config.spec_unknown_key`, its message naming the key; on the run's terminal `detail.keys` reads `«redacted»`, the terminal's reading of that member for every refused key — `[redacted]` in 7.67.1, the marks of #3970 below) before any model turn; a resume whose re-supplied config writes it is refused at the same door, as `checkpoint.config_refused` carrying that code (the door's existing resume face, `test/backlog1950-resume-config-doors-preclaim.test.ts`). The door judges own keys by name, so `false` is refused like `true`; a value of the name that a spec only inherits is neither read nor refused, the door's law for every inherited name (it was read before, and changed nothing).
|
|
715
|
+
- **Hard-breaking — one, three sentences (open to veto @server @cli @client-core).** *What changed:* `TaskSpec.streamingToolExecution` is removed from the type and refused by name at run time. *Who is affected:* a caller whose spec literal names the key (a compile error) or who hands it in at run time (the run is refused, where it used to run exactly as it would have without the key), and a host that mirrors `keyof TaskSpec` in a key list of its own. *What to do:* drop the key — a Runner run is the same without it — and remove it from any mirrored key list; a workflow resumed over a journal whose agent calls wrote the key replays them only while its script still writes it (the first call it runs live is then refused), and dropping the key changes those calls' identity, so they run live again from the first one (the journal's prefix rule).
|
|
716
|
+
- **Not changed.** The order a Runner leg runs tool calls in (after the stream final, batched by declared safety under the concurrency cap).
|
|
717
|
+
- **Divergence (D-3830-a, open to veto).** The reference starts every streamed tool call in-stream and runs its permission check per call mid-stream: every streamed `tool_use` block is handed to its streaming executor (`cli316.js:278219`), which carries the permission callback into each call's run (`cli316.js:278198`). A sema Runner leg arms a gate on every leg, and a gate that may park an ask needs the assistant message persisted first; starting calls mid-stream needs a gate that judges a call mid-stream without parking it, which does not exist. The divergence held since the integrity guards became always-on; this release states it instead of keeping a switch that read as the opposite.
|
|
718
|
+
- **Rules.** Retired: the task member, the harness option and the harness's forwarding predicate (opt-in and no `tool_call` handler), the docstring's "no effect on a gated task" exception, the liveness table's row (it was proved by a loop-level test, never a Runner run), and the crash-window entry in `docs/KNOWN-ISSUES.md` (no Runner run reaches that window). Added: none — the refusal is the member door's.
|
|
719
|
+
- **Pins.** `test/backlog3830-streaming-tool-execution-retired.test.ts` — R1 a spec writing the key is refused `config.spec_unknown_key` naming it, zero turns, no tool runs (and no call ever starts before its stream final) · R2 the member is gone from `TaskSpec` and the member table, the option from `AgentHarnessOptions` (type-level `@ts-expect-error` both) · H1 a harness handed the key reads nothing: with no `tool_call` handler, the safe call starts after the stream final · P0 control: a Runner run without the key runs the safe call after the stream final and completes. Baseline (`origin/stage-7670` `cb66f49bf`) red on R1 / R2 / H1 — R1 red only at the refusal: the run with the key completed and no call started before its stream final, the measured form of "never took effect" — the control green; green here. Re-pinned: `test/streaming-tool-execution.test.ts` (its harness-snapshot cell is H1 now; the file itself goes with the loop lane, #4032 below), `test/public-knob-liveness.test.ts` (the row). Mutations M1–M3 each red: M1 the member back in the table (R1, R2), M2 the harness reading and forwarding the key again when no `tool_call` handler is registered (H1), M3 the member door letting an unknown own key through (R1). (A fourth control — the loop's own lane still starting the fixture's call mid-stream — went with that lane, #4032 below.)
|
|
720
|
+
- **Consumers.** @server: `src/trace/core-keyset-guard.ts` lists the key in `TaskSpecUnwired` (zero write points) — delete the word and regenerate `docs/core-consumption/READ-TABLE.json`; nothing changes at run time. @cli: nothing to read (the shell's own `streamingToolExecution` is its query loop's gate, not this member). @client-core: nothing to read (one comment in `src/adapt/textStream.ts` names the term; no code reads the member).
|
|
721
|
+
|
|
722
|
+
### Where a finished workflow's full record is has one answer — `WorkflowRun.recordArtifact`, written by the engine in the run's terminal transition and read through one root export, `workflowRecordPickupOf` (#3980; ruled C-R1321; behavior change: with a result store wired, `run_end` and the terminal write wait for the record write, bounded at 10 s; narrowing: none; widening: none; hard-breaking: none; closed sets: none registered — the member's `kind` is three words, `stored` / `unconfirmed` / `no_store`; exports: +1 `workflowRecordPickupOf`; wire: `WorkflowRun.recordArtifact?`, `RunWorkflowOptions.taskArtifacts?`, additive) (@server @client-core @sdk @test)
|
|
723
|
+
|
|
724
|
+
- **Rule.** A workflow run's record carries one member, `recordArtifact`, that says where its whole record (the result and every agent's row) can be read back: `{ kind: "stored", ref }`, `{ kind: "unconfirmed" }` or `{ kind: "no_store" }`. The engine writes it once, in the run's one terminal transition: it renders and stores the record through `RunWorkflowOptions.taskArtifacts`, sets the member, then emits `run_end`, and the member lands in the same terminal write as `status`. Every face that names it to a model decodes and renders it through `workflowRecordPickupOf(raw)` — the one decode (taken through the engine's foreign-value receipt: own data members only, no accessor or Proxy trap runs; `ref` present iff the `stored` shape with a ref of the shape the result store mints for a task's result; anything else reads as absent) and the one sentence. A finalized run is written by its terminal only: a park answered after the run ended, that park's own row write, and the retry bookkeeping of a call still in flight write nothing on it.
|
|
725
|
+
- **Before.** The Workflow tool stored the record after the run had settled and named the ref only inside the completion frame's `diagnostics` text (display-only, never handed back to a model by a host); the run record — the one thing a read face and a polling host can reach — said nothing about it, and the terminal was written before the record was stored, so a process that died between the two left the pickup nowhere.
|
|
726
|
+
- **Now.** The member is on every terminal record this engine writes (absent while running, on an older engine's record, on a row a deployment's sweep closed, and on a terminal write that did not land — so absence on a terminal record is not `no_store`). The record write is bounded by `WORKFLOW_STORE_WAIT_BOUND_MS` (10 s) on a timer that holds the process; past the bound, or when the render, the ref mint or the write throws, the member is `unconfirmed`, one run-log line names the bound, and a late landing is never written back. The completion frame's `diagnostics` sentence is the decoder's `line`, byte for byte the sentence it was. `stored` promises the write landed at the terminal, not that the deployment's result store still keeps it. The decoder judges a `stored` ref segment by segment (#3988) with the result store's own segment test — four segments (`tr_<session>`, `fold.task`, the task's, the content coordinate), each non-empty, in the ref charset and within the segment ceiling — so a ref carrying a path separator, a control character, a quote or whitespace, an empty or over-long segment, or another segment count reads as the record unread and is never echoed into the sentence or the returned `ref` (a narrowing, open to veto @client-core @server: every ref the engine mints reads as before; pins `test/backlog3988-workflow-record-ref-segments.test.ts`).
|
|
727
|
+
- **Not changed.** The completion frames (`task_notification`, `workflow_complete`) and the notifier's input gain no key; `diagnostics` reads as before; the list summary (`WorkflowRunSummary`) and the poll projection carry no new key; a background agent's result keeps its own record (`finalOutput` / `finalOutputFull`). The module-internal bound constant `WORKFLOW_RESUME_CLAIM_FINALIZE_TIMEOUT_MS` is renamed `WORKFLOW_STORE_WAIT_BOUND_MS` (not a root export; it bounds the resume reads, the park write and the claim release as before). Residual: a process that dies inside the one record write leaves the row `running` (KL-3960-b).
|
|
728
|
+
- **Divergence (D-3960-a, ruled C-R1321, open to veto).** The reference carries a structured pickup on the completion frame itself (its SDK task-notification frame's output file, taken from the task row); this engine keeps the pickup on the run record only, and the completion frames carry none — once the originating leg is gone the model's completion notice is built by the host from structured facts, and `diagnostics` is display text a host never hands back to a model.
|
|
729
|
+
- **Rules.** Added: one record member (the fact itself), one engine seat, one holding position on the existing terminal bound, one decode-and-render export. Retired: the tool's own write of the record (it reads the member), and the sentence it worded in its own module (now the decoder's). No special case: the record step is one helper both terminal paths call.
|
|
730
|
+
- **Pins.** `test/backlog3980-workflow-record-artifact.test.ts` — G1 / G2 / G3 the first terminal write carries `stored` (completed, failed, parked; a resume gets its own) · G4 one ref on the frame, the record and the task list · G5 / G12 a throwing write or seat reads `unconfirmed` and the terminal still lands · G6 / G13 a write that never settles, or lands late, stays `unconfirmed` · G7 `no_store` · G8 never on a running record · G9 never on the frames (negative control) · G10 the Runner's own Workflow mount · G11 in-memory and file round trip · G14 a resumed run · G-ord record < `run_end` < terminal write · G-snap the stored record is the durable one · G-clock a clock that throws while the bound is named reads `unconfirmed`, no second terminal · G-snap-park a park answered after the run ended writes nothing on it · G-timeout / G-cancel · G-norunstore · G-K11 the decoder (three readings, twenty malformed values incl. a ref not of the minted shape and a valid value behind a Proxy, no accessor or trap run) · G-hold the bound holds the process (child process) · G-type; `test/backlog3980-workflow-record-artifact-seat.test.ts` G-seat (no seat, no render).
|
|
731
|
+
- **Consumers.** @server: the read face's `WORKFLOW_RUN_DISPOSITIONS` gets two cells — `recordArtifact: { nested: { kind: "forward", ref: "forward" } }` (the projection vocabulary can also spell `strip`; only `nested` keeps the fact) and `completionId: "strip"` → `"forward"`; the keyset guard reds until both are written. @client-core: a pushed completion reads the run record once before delivery (bounded, delivered once) and renders the pickup through `workflowRecordPickupOf`; the polling arm takes it from the record it already reads. @sdk: `WorkflowRun` gains `recordArtifact?` (and `completionId?` on the read face). @test: HB-1–HB-9 in `docs/release-criteria/7.68.0.json` (HB-PUSH / HB-POLL / HB-PUSH-RF / HB-PUSH-UC wait on client-core).
|
|
732
|
+
|
|
733
|
+
### The Read family reads the object it judged — the Read tool, a compaction's restore, the write-target card's preview and the Artifact tool's local file open the judged key through the one entry (`readObjectAt`): held where the environment binds, else the one by-name simulation; a held PDF reaches poppler as bytes on its standard input (`ExecutionEnvExecOptions.stdinFrom`), never as its path (#4036 = #4034 S1c; ruled C-R1357 with notes 1–3; behavior change: a link put on the way of a judged key before the read is refused where it was read through, a held PDF is read in full where a carrier's held object made it a reference, an Artifact file is never read past its limit; narrowing: one, below, open to veto @cli @test; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: `ExecutionEnvExecOptions.stdinFrom?` and `ReadObject.byName?`, additive) (@cli @server @test)
|
|
734
|
+
|
|
735
|
+
- **Rule.** Every byte, size and kind a Read-family read uses comes from one object opened at the judged key (`readObjectAt`, `src/tools/fs/bound-open.ts`): `bound` ⇒ the held object; `unbindable` (no `openBound`, or `not_supported`) ⇒ the one by-name simulation, judged again by the read's own resolution asked again (the path, on the same tables, still resolves to the judged key — an environment whose resolver judges by spelling judges again by spelling); refused ⇒ the read's own sentence (`precondition_failed` ⇒ the moved file's "resolved to a different file while it was being checked"). A program reads a held object one way: its bytes on standard input (`-`); a simulated object's programs read the key by name, judged again right before each start. Only `pdfAsReference` makes a PDF a reference.
|
|
736
|
+
- **Before.** The Read tool, a compaction's restore and the card preview resolved the path and then read it by name (`fileInfo`, `readBinaryFile`); an ancestor turned into a link right after the judgment had them read the file it led to — outside the roots or under a read-deny row (a real race: 44 of 2000 reads of a flipped ancestor returned the outside file). poppler (`pdfinfo` / `pdftoppm` / `pdftotext`) was started on the key by name. A carrier's held object made its PDF a reference whether or not it asked. The Artifact tool's local file was read whole and only then held to the host's limit.
|
|
737
|
+
- **Now.** The same flip: the held open refuses it (darwin: the open refuses a link on the way; Linux: the descriptor's path is not the key) and the simulation's judgment asked again refuses it — the moved file's refusal, no byte of the other file (`readText`: `moved`). A held PDF's programs read the held bytes on standard input; a held PDF read in full is the document. The Artifact tool's file is held to the limit by its size before the read and by the read itself (never more than the limit taken; one grown past it while read is `too_large`, `judgedOn: "bytes"`). A held object whose size cannot be read is that object's answer: the path is not asked again (no ENOENT correction from the path after a held open).
|
|
738
|
+
- **Narrowing (open to veto @cli @test).** On a platform that binds (darwin whose kernel refuses links on the way, Linux with `/proc`) a Read whose judged key is reached through a link put on its way after the judgment is refused with the moved file's sentence (`details.type: "target_moved"`), where it read the file the link led to; the card preview answers `moved`.
|
|
739
|
+
- **Restored, not widened.** A PDF read from a carrier's held object without `pdfAsReference` is the document again (the reference was a side effect of holding). A person's PDF attachment still asks for the reference explicitly (`pdfAsReference: true`): its size and page count, no program started.
|
|
740
|
+
- **New optional member (@server).** `ExecutionEnvExecOptions.stdinFrom?: Uint8Array` — the bytes a start's standard input carries, then its end (`NodeExecutionEnv` writes them; `spawnBackground` takes no input: `BackgroundSpawnOptions` omits it). It reaches only an environment whose `openBound` held the file: an adapter that implements `openBound` writes it, or its PDF reads fail with poppler's own error (empty input) — never by reading the path. A poppler that cannot read `-` fails the same way (documented for poppler 22.12; not retried on the path). Integration guide §11.4. `ReadObject.byName?` (reached through `LspReadAdmission.object`) — present on the by-name simulation only: its key and its judgment asked again, which the PDF lane asks right before a program starts on the key; an LSP read gate that builds its own `ReadObject` leaves it out (a held object has none).
|
|
741
|
+
- **Not changed.** A path attachment (bound or `cannot_bind`); the plan file's and the LSP session's readers (this section leaves them; #4034 S1a, below, moves the LSP session's reads onto the simulation); the Read on an environment that cannot bind reads as it did, byte for byte, where nothing moves (the simulation's extra looks are the cost); the wire otherwise.
|
|
742
|
+
- **Also.** A held PDF's bytes are read once (up to the 100 MB extraction cap) and handed to each program — the reference's form off Linux; a by-name PDF grown past the 20 MB whole-send size while it is read is refused naming the cap, where it was read on by name; a held PDF grown since its size check is judged on the bytes read (over 20 MB ⇒ not sent whole). A FIFO at the key: the Read's sentence names the opening's answer. An opening the environment refuses (the mode forbids reading, descriptors exhausted, a busy or failing device) is a failed read — `cannot read …: <the open's reason>`, the sentence the by-name read gave — and only a kind and size that could not be taken (a look that failed, a held object whose size cannot be read) is the type / stat sentence; the Artifact tool says `cannot read` / `cannot stat` by the same step.
|
|
743
|
+
- **Known limits.** KL-1503-b now covers the Read family (the simulation's window; a simulated PDF's programs judged again before each start).
|
|
744
|
+
- **Rules.** Retired: `readForModel`'s two read arms (bound / by path) — one object; `readText`'s own by-name read; the "held ⇒ PDF reference" coupling; the Artifact tool's unbounded read; the Read's three missing-file refusals (image, PDF, text lanes) and the PDF reader's own copy — one (`missing`, asked only when the opening found nothing). Added: the one program-input form (`PdfInput`: held bytes on standard input, or the simulation's key judged right before the start); `readWithin` (a whole read under a limit, its size judged before and after) shared by the card preview and the Artifact tool; `sameKeyAgain` (a reader's judgment asked again: its own resolution answers the same key) shared by the Read family and the Artifact tool.
|
|
745
|
+
- **Pins.** `test/backlog4036-read-bound.test.ts` R1–R15 — R1 / R2 the flip after the judgment, bound and simulated ⇒ moved, no outside or read-denied byte · R2b nothing moved ⇒ bound and simulated answer alike · R2c the judgment asked again is the read's own resolution (a leaf the environment cannot canonicalize, judged through its ancestors, is still read) · R3 / R3c a held PDF's pages / text layer: poppler reads `-`, the held bytes, while the key is a link outside · R3b simulated: the key turned into a link while the page count runs ⇒ the renderer is not started · R4 a held PDF read in full is the document · R5 a person's PDF attachment is a reference, no program started · R6 the Artifact file grown past the limit ⇒ `payload_too_large`, every by-name read limited · R6b the Artifact file under the same flip ⇒ not read (`precondition_failed`), bound or simulated · R7 the card preview ⇒ `moved` · R8 `stdinFrom` carries its bytes (twice, 1 MiB) · R9 why bytes: a descriptor handed to one start is consumed · R10 the deployed poppler reads `-` (where poppler is installed) · R11 a real race, 2000 reads ⇒ 0 outside · R12 / R13 a refused opening (a file whose mode forbids reading; permission, descriptors, busy, i/o, name errors) ⇒ `cannot read …: <reason>` in every lane, bound and simulated alike; a look or a size that fails keeps the stat sentence · R14 a FIFO is still refused unwaited, its words unchanged · R15 the Artifact tool's `cannot read` / `cannot stat`. Re-pinned: `test/pdf-exec-error-truncation.test.ts` and `test/backlog1795-time-cut-consumers.test.ts` (the PDF legs take a `PdfInput`), `test/backlog3609-bounded-read-primitive.test.ts` R1–R5 (its path-reading environment has no `openBound`; its grown file grows as it is read; R5: a by-name PDF grown past the whole-send size is refused naming the cap).
|
|
746
|
+
- **Consumers.** @cli: macOS / Linux hosts — a Read through a link swapped in after the judgment is refused (`target_moved`). @server: adapters that implement `openBound` write `stdinFrom` (§11.4); an adapter without `openBound` reads through the by-name simulation, the same bytes where nothing moves but more questions to the environment — 3 `fileInfo` and 4 `canonicalPath` calls per Read where 7.67.1 made 1 and 1 (measured by the release review; KL-4036-a, #4272) — on a remote environment each is a round trip (inferred; please measure). @test: a Read flip cell ⇒ 0 outside bytes; a held PDF's pages read with poppler on standard input.
|
|
747
|
+
|
|
748
|
+
### A platform that cannot place an opened object by one kernel observation answers `not_supported` (win32, the BSDs); `setFileMode` changes a mode through that same placed open; a reader whose no-binding disposition is to read by name — the plan file's reader and now the LSP session's reads — reads ONE simulation of the bound object (#4034 S1a; ruled C-R1357 with notes 1–3; behavior change: win32 / BSD path attachments, `setFileMode` through a link or onto a FIFO, LSP reads on an environment that cannot bind, a bounded by-name read of a FIFO; narrowing: two, below, open to veto @cli @server; widening: one, below, open to veto @server; hard-breaking: one, type level, below, open to veto @server @cli; closed sets: none; exports: none; wire: none) (@cli @server @test)
|
|
749
|
+
|
|
750
|
+
- **Rule.** Where an opened object is, is one kernel observation or nothing: darwin's open (`O_NOFOLLOW_ANY`, #4034 S0), Linux's descriptor path (`/proc/self/fd/<fd>`); a platform with neither answers `not_supported` and opens nothing. `NodeExecutionEnv.openBound` and `setFileMode` take that one placed open (read-only, non-blocking, a regular file or a directory). A reader that may read what it judged by name declares it once (`readObjectAt` with its own re-judgment, `src/tools/fs/bound-open.ts`): where the environment binds it reads the bound object, where it cannot (no `openBound`, or `not_supported`) it reads the one simulation — every answer (kind, size, bytes) comes from a look at the leaf without following a link followed by the reader's judgment asked again, and the bytes from one bounded read of the key (`readBinaryFile(key, signal, max + 1)`) with the look and the judgment after it (KL-1503-b). A person's path attachment does not read by name: it is bound or not attached.
|
|
751
|
+
- **Before.** win32 and the BSDs placed the opened object by reading the path twice (its `lstat` identity, then its canonical spelling); an ancestor turned into a link for the open and put back between the two reads passed. `setFileMode` opened with only the leaf unfollowed and without `O_NONBLOCK`: an ancestor that was a link re-permissioned the file behind it, and a FIFO swapped in at the key waited for a writer. The LSP session gate refused every read on an environment that could not bind ("this environment cannot bind a read to the file that was checked"), its target facts said nothing there, and the plan file's by-name arm read the whole file with an unbounded, blocking read. The environment's bounded by-name read (`readBinaryFile(path, signal, maxBytes)`) opened as `readFile` does: a FIFO at the path waited for a writer. `openBound` asked whether the call was aborted only before the kernel probe.
|
|
752
|
+
- **Now.** win32 / BSD: `openBound` and `setFileMode` answer `not_supported` ("This platform gives no kernel observation of where an opened object is, …"). `setFileMode`: a link anywhere on the path ⇒ `precondition_failed` (darwin: the open; Linux: the descriptor's path is not the key), a FIFO ⇒ `invalid` at once. The LSP gate and its target facts read the simulation where the environment cannot bind: the file is read, and a link put on its way before the by-name read is caught by the judgment after it (`precondition_failed`, no byte). The bounded by-name read opens without blocking and reads a regular file only (an opened FIFO or device `invalid`, a directory `is_directory` as before; a socket keeps the kernel's refusal). A cancelled signal answers `aborted` after the probe, after the placement (no mode is changed) and in the simulation's judgment (not a changed file).
|
|
753
|
+
- **Narrowing 1 (open to veto @cli).** On win32 and the BSDs every path attachment is not attached (`cannot_bind`), where the two path reads attached it (C-R1357 Q2). Host-held text and images are unaffected. win32 LSP reads are not narrowed: they read the simulation (before: the two path reads held the file and it was read).
|
|
754
|
+
- **Widening (open to veto @server).** An environment with no `openBound` — a remote lane — has its LSP session's reads read through the simulation, where they were refused in one sentence; the same posture as its plan file reader. Its LSP target facts ("File does not exist", "Path is not a file", the size cap) now come from the simulation's look.
|
|
755
|
+
- **Narrowing 2 (open to veto @cli @server; the link half is a defect the contract already ruled out).** `setFileMode` on a path that is not canonical (a link anywhere on it) is refused rather than following the link: a rewind restoring a recorded mode through such a path reports it as not preserved (`FileHistoryRestoreResult.modeNotPreserved`). The engine's own restore keys are canonical. Where the placed open answers `not_supported` (the BSDs, a Linux host without `/proc`, a macOS whose kernel does not refuse links on the way) a recorded mode is likewise reported not preserved, where it was applied through the leaf-only open. The `FileSystem.setFileMode` contract now says "no symlink on the path followed, leaf or directory".
|
|
756
|
+
- **Not changed.** Linux and darwin `openBound` answers; the attachment reader (bound or `cannot_bind`); `listOpenDirectory` on darwin (KL-1311a2-g, its sentence made true: the entries are judged as names under the held directory's spelling, and `list()` sizes / times are the listed entries'); the Read tool's own by-name read (#4036); the wire.
|
|
757
|
+
- **Hard-breaking — the LSP read gate admits a read object (type level; open to veto @server @cli).** *What changed:* `LspReadAdmission`'s admitted arm (`ok: true`) carries `object: ReadObject` — `kind`, `size`, `readBytes`, `close` and an optional `byName` — where it carried the whole `BoundFileObject` (the one a session reads may be the by-name simulation, which has no held identity or listing). *Who is affected:* a host that implements `LspSession` itself and reads a member of the admitted `object` beyond those (`identity`, `list`): it no longer compiles; a host that writes an `LspReadGate` and admits a `BoundFileObject` still compiles, and the engine's own `TransportLspSession` reads the object through its read members and `close` only. *What to do:* read the file through `readBytes` / `size` and close the object, as the bundled session does.
|
|
758
|
+
- **Known limits.** KL-1311a2-h retired (no path-check arm remains). KL-1311a2-g rewritten. KL-1503-b now covers the LSP readers too. New KL-4034-a: a bound read binds the path's placement, not every name of the object (a hard link or a mount point inside the roots reads what it names; read-deny rows are path rows) — the reference's shape; DEFERRED security-low (the model cannot make a link: every link-making command is refused at the read station).
|
|
759
|
+
- **Rules.** Retired: the non-Linux, non-darwin path check (two observations), `setFileMode`'s own open, the LSP gate's refusal sentence, the plan reader's inline by-name arm, win32's directory-listing refusal arm. Added: one declaration of "this reader may read by name" (`readObjectAt`), which every such reader passes its re-judgment to; the simulation is one object (`simulatedObjectAt`) the plan reader and the LSP readers share.
|
|
760
|
+
- **Pins.** `test/backlog4034-one-observation.test.ts` S1a-1 … S1a-9 (14 cells) — S1a-1 win32 / freebsd ⇒ `not_supported`, nothing opened · S1a-2a / S1a-2b `setFileMode` with a link on the way ⇒ refused, the file behind it keeps its mode, a plain file still gets its mode (darwin real kernel; Linux arm) · S1a-3 `setFileMode` on a FIFO ⇒ `invalid` within 3 s · S1a-3b a signal cancelled during the placement ⇒ `aborted`, mode untouched · S1a-4 the LSP gate without `openBound` admits and reads the file; a link put on the way before the by-name read ⇒ `precondition_failed`, no secret byte · S1a-4b LSP target facts from the simulation · S1a-4c a directory re-pointed outside between the tool's judgment and the facts' look ⇒ no size class, no existence said · S1a-4d a cancelled re-judgment ⇒ `aborted` · S1a-5 a FIFO at the key of a bounded by-name read, and swapped in for the plan file after its look ⇒ refused within 3 s · S1a-6 a file over the cap swapped in after the look ⇒ refused, the by-name read answers at most the cap + 1 · S1a-7 an abort that lands during the probe ⇒ `aborted`, nothing opened · S1a-8 the known limits / guide sentences · S1a-9 one simulation (no reader reads the judged key by name itself). Mutations M9–M16b and M3x, each red on its cells (the test file's header). Re-pinned: `test/backlog4035-not-supported-one-reading.test.ts` G3a / G3b (absent and `not_supported` read the one simulation alike), `test/backlog1293c-lsp-bound-read.test.ts` M21 (an environment without `openBound`: the session reads the judged file through the simulation, never through the deployment's `readText`), `test/backlog1701-win32-mode-bits.test.ts` (the POSIX half loads under the host's own platform, over a canonical path), `test/backlog3609-bounded-read-primitive.test.ts` N7 (a device at the path is refused; a procfs file is the size-less stream, Linux), `test/backlog3757-3758-adapter-message-reading.test.ts` (plan reader splices 9 → 7), `test/file-history-store.test.ts` and the file-history contract kit (mode cells over a canonical root, as the engine's restore root is).
|
|
761
|
+
- **Consumers.** @cli: win32 / BSD hosts — path attachments `cannot_bind`; LSP reads unchanged. @server: lanes without `openBound` — LSP reads and target facts through the simulation; S-3356's own adapters (answering `not_supported` on a non-Linux target) are read the same way. @test: a win32 attachment cell ⇒ `cannot_bind`; an LSP cell on a lane without `openBound` ⇒ the file's content.
|
|
762
|
+
|
|
763
|
+
### On macOS a bound open is one observation — `NodeExecutionEnv.openBound` opens with `O_NOFOLLOW_ANY`, so the open itself refuses a link on any component of the path, and a kernel that does not refuse them answers `not_supported` (#4034 S0; ruled C-R1357; behavior change: darwin only; narrowing: one, below, open to veto @cli @server; widening: one, below, open to veto @cli @server; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @server @test)
|
|
764
|
+
|
|
765
|
+
- **Rule.** Where the held object is, is one kernel observation: on darwin the open (`O_NOFOLLOW_ANY`, `0x20000000`, in place of `O_NOFOLLOW` — the kernel refuses the two together), on Linux the descriptor's path (`/proc/self/fd/<fd>`, unchanged). Whether this kernel refuses a link on the way is asked once per process, of the link every macOS has: an open of `/var/tmp` under the flag must answer `ELOOP` where `/var/tmp` resolves without it (`stat`: the `ELOOP` is then the flag's, not a loop of the layout's own, and a host that may read `/var/tmp`'s metadata but not the directory can still ask).
|
|
766
|
+
- **Before.** darwin opened without following the leaf only, then read the path twice (its `lstat` identity against the descriptor's `fstat`, then its canonical spelling): an ancestor turned into a link for the open and the identity read, and put back before the canonical read, passed — `openBound` held a file outside the roots and a person's path attachment carried its bytes (`attached`).
|
|
767
|
+
- **Now.** That shape is refused at the open (`ELOOP` ⇒ `precondition_failed`, "A link stands on the path now, so it is not the object that was checked" — the same sentence on every platform); the path is not read after the open.
|
|
768
|
+
- **Narrowing (open to veto @cli @server).** A macOS kernel that walks through a link under the flag, rejects it (`EINVAL`), or a probe that cannot ask (`/var` not a link, `/var/tmp` not resolving, any other answer — asked again by the next open) answers `not_supported`: a path attachment is `cannot_bind`; the plan file reader and the LSP session read the one by-name simulation there (`not_supported` reads as the member's absence, #4035, and the model's own reads take the simulation, #4034 S1a — both ship with this change). The flag is macOS 11's; an older kernel is answered by the probe, never assumed.
|
|
769
|
+
- **Widening (open to veto @cli @server).** A file renamed or replaced at the key in the instant after the open is read as the object the open placed (it was `precondition_failed`): that object was at the judged key, through no link, when it was opened.
|
|
770
|
+
- **Not changed.** Linux; `listOpenDirectory` (KL-1311a2-g); the `openBound` contract text (#4035). win32, the other platforms and `setFileMode` change in the same version (#4034 S1a).
|
|
771
|
+
- **Rules.** Retired on darwin: the two path reads after the open. Added: one capability probe (the open is one observation only where the kernel refuses links on the way).
|
|
772
|
+
- **Pins.** `test/backlog4034-one-observation.test.ts` — C0 control · C1 an ancestor that is a link for the open, put back before the canonical read ⇒ refused · C2 the person's attachment of it ⇒ not attached, no byte · C3 put back after ⇒ refused · C4a / C4b / C4c / C4d a kernel that ignores / rejects the flag, a probe that cannot ask, a layout whose `/var/tmp` is itself a loop ⇒ `not_supported` (C4c asked again; C4d: a host that may not read `/var/tmp` still asks and holds) · C5 a plain file, a directory and the attachment are held and read, the open asks `O_NOFOLLOW_ANY` and not `O_NOFOLLOW` · C6 the file replaced right after the open ⇒ the object the open placed is read · C7 a FIFO swapped in at the judged key ⇒ `invalid` without waiting for a writer, the attachment not attached. Mutations: M1 `leaf-only` (C1 C2 C3 C5 red) · M2 `probe-fail-open` (C4 red) · M3 `second-look` (C6 red) · M4 `no-control` (C4d red) · M5 `memo-forever` (C4c red) · M6 `eloop-fallback` (C1 C2 red) · M7 `nonblock-darwin` (C7 red) · M8 `control-open` (C4d red).
|
|
773
|
+
- **Consumers.** @cli: macOS host attachments — the swap shape is now `not_attached`; on a macOS whose kernel does not refuse links on the way every path attachment is `cannot_bind`. @server: Linux lanes read the same; a macOS device executor built on `NodeExecutionEnv.openBound` gets the darwin arm. @test: darwin black-box cell — the R-B4 attachment shape (`attached` with the outside file's bytes) ⇒ `not_attached`.
|
|
774
|
+
|
|
775
|
+
### A run that completes strictly past a ceiling it was given says which, how far and by how much — `TaskResult.limitCrossed` and the operator notice `task.completed_past_limit` (#4184; test c767a; behavior change: one optional result member and one operator notice, the verdict and every default reading are unchanged; narrowing: none; widening: none; hard-breaking: one notice word, type level only (below); closed sets: `EngineNoticeCode` +1; exports: none added, none removed; wire: one optional result member and one notice code, additive) (@server @cli @client-core @test)
|
|
776
|
+
|
|
777
|
+
- **Before.** A run whose last turn answered and owed no further request completed past its declared ceiling with nothing said on a resource-slice run: a 500-token ceiling against `stats.tokens` 1000 was a bare `completed` (test c767a G6.f), the caller compared the numbers itself, and the walltime window had no figure on the result at all. (Measured on the stage base: a plain run that crosses any of the four axes with a text answer already completes with `task.interrupt_unconsumed`, audience `user`, whose `origin` is `budget`, `turns` or `walltime` — a word and no figure; a resource-slice run crossing tokens, turns or walltime completes with no notice and no member.)
|
|
778
|
+
- **Now.** One evaluation at the run's settle (`limitCrossedOf`, `src/core/runner/clock-and-limits.ts`) feeds both faces. `TaskResult.limitCrossed` — `{ axis, limit, observed }`, present only on a `completed` terminal that went **strictly** past a ceiling — and one `task.completed_past_limit` notice (audience `operator`, once per run, `detail: { axis, limit, observed, sessionId, runId, taskId }`). `axis` is a parked run's `resource_limit` word (`budget_tokens`, `budget_cost`, `walltime`); `limit` is the ceiling the engine compared (for tokens on a resource-slice run, the smaller of the slice window and what the cross-slice allocation had left) and `observed` the run's figure, in the axis's unit: tokens, **micro-USD** (`stats.costMicroUsd`'s), **milliseconds** (the walltime figure is what the turn boundary read when it found the window passed, recorded once on the run state, so work after the answer does not move it; when the walltime timer fired and no boundary read the window, the figure is the run's elapsed milliseconds at its settle — `limitCrossedOf`, `src/core/runner/clock-and-limits.ts` — which does include that work). One axis: the first of tokens, cost, walltime among those the engine recorded. A turn that reaches the turn cap and also crosses a ceiling still reports the ceiling (the cap's handler now reads the budget check where it records the cap; the failure code of a run that owed a turn is not touched). The sentence says why the verdict stands: the last turn answered and owed no further request, so nothing was cut.
|
|
779
|
+
- **Strictly past, and no `turns` word.** A ceiling met exactly says nothing, and so does a run that answers on the turn its cap allows (`maxTurns: 1` answered in one turn, `maxTurns: 3` answered on the third): that is an allowance used, not crossed — a completed run is never past its turn cap, because reaching it with a turn owed fails or parks. (A first cut reported the cap as `turns`; every one-question run under `maxTurns: 1` then carried a false "crossed" fact. The axis type has three words.) The stop notice's own `turns` origin on a plain run that reaches the cap is the existing form and is not changed here.
|
|
780
|
+
- **On a plain run the two notices coexist (conditionally accepted).** `task.interrupt_unconsumed` says a stop landed after the answer (audience `user`, `origin` from its own order: stop verb, timer, turn cap, budget gate); the new one says which ceiling and how far (audience `operator`). Same fact, two questions. The words line up as `budget_tokens` / `budget_cost` → `budget`, `walltime` → `walltime` (pinned). Measured, when two ceilings are crossed the two orders can name different axes: tokens + walltime on a plain run reads `limitCrossed.axis` `budget_tokens` beside `origin` `walltime`, and tokens + the turn cap on one turn reads `budget_tokens` beside `origin` `turns`. Not changed here; folding the stop notice's limit origins into this word is a separate ticket.
|
|
781
|
+
- **Not changed.** The verdict (`completed`), `stats`, the failure of a run that owed another request (`limits.max_*_exceeded`), the park of a resource-slice run that owed one, `task.interrupt_unconsumed` (words, audience, firing), the order the park attributes in.
|
|
782
|
+
- **Hard-breaking (type level only).** (1) *What changed:* `EngineNoticeCode` gains `task.completed_past_limit` (`ENGINE_NOTICE_CODES` +1; the release's count, 135 → 153, is in the #3832 section). *Who is hurt:* a consumer that keys a record by `EngineNoticeCode` — a notice allowlist or router, an audience table, a renderer — gets a compile error until it gives the word a row. The wire is additive (`TaskResult` gains one optional key; a closed set that gained a member). *Migration:* add the row (operator audience; branch on `detail.axis`, never on the sentence); read `limitCrossed` where the figures are wanted.
|
|
783
|
+
- **Rules.** Readings of "which ceiling did this run end past" 1 (the stop notice's `origin` chain) → 2 (that chain, untouched and pinned against the new one, and `limitCrossedOf`, which the result member and the new notice share — one evaluation for the two of them); the second reading goes when the stop notice's limit origins fold into the new word (a separate ticket). The walltime window's crossing is read at the boundary that already read it, and recorded once (`RunState.limits.walltimeCrossedMs`); the budget check a turn-cap turn skips is read once where the cap is recorded (`RunState.limits.budgetOverAtTurnCap`). Added: one result member, one notice code, two run-state members. Retired: none.
|
|
784
|
+
- **Closed sets.** `EngineNoticeCode` +1 (three tables in `src/core/governance-codes.ts`; docs/CLOSED-SETS.md). The word is in the `task.*` family beside `task.interrupt_unconsumed` (a sentence the run says at its ending), not under the `limits.*` prefix a terminal failure code carries.
|
|
785
|
+
- **Consumers.** @server @cli @client-core: a row for `task.completed_past_limit` in the notice table (operator audience; do not render it to the person talking to the run: they got their answer); read `TaskResult.limitCrossed` for the machine-readable fact; the units differ by axis (tokens, micro-USD, ms). @test: c767a G6.f re-read — the 500-token run reads `limitCrossed { budget_tokens, 500, 1000 }` and one operator notice; a ceiling met exactly, the turn cap reached on the answer and the unchanged controls read none.
|
|
786
|
+
- **Pins.** `test/backlog4184-completed-past-axis.test.ts` — A1 / A2 / A4 a resource-slice run completes past tokens (500 / 1000), cost (5000 / 10000 micro-USD), walltime (60 / ≥ 150 ms): the member, one operator notice, the same facts · A3 / A3b the turn cap reached on the answer (`maxTurns` 1; 3 with two tool turns before) ⇒ no member, no notice · A5 the ceiling reported is the one compared (a 500-token window with 300 left on the cross-slice allocation reads 300) · A6 a ceiling met exactly ⇒ none, one over ⇒ reported, on a slice run and a plain run · B-tokens / B-cost / B-walltime a plain run: both notices, the stop notice's origin is the table's word for the axis · B-turns a plain run that answers on its cap ⇒ no member, no new notice · B-multi-a / B-multi-b the two measured multi-ceiling readings above · D1 / D2 `limitCrossedOf` read directly (equal ⇒ nothing, one over ⇒ reported, the order, a precall hit and an unset ceiling) · C1 every axis declared and none crossed ⇒ no notice, the key absent, plain and slice · C2 no limits ⇒ none · C3 / C4 a run that owed a turn (plain failed, slice paused) carries neither · C5 the default notice seat gets it once. Control: all green with the 3873 and 4072 pins. Mutations, per fixture: the member on every terminal ⇒ C3 red · the notice removed ⇒ A1 A2 A4 A6 B-tokens B-cost B-walltime B-multi-a C5 red · the member removed ⇒ A1 A2 A4 A5 A6 B-tokens B-cost B-walltime B-multi-a B-multi-b C5 red · walltime checked before tokens ⇒ B-multi-b D2 red · the limit read as the figure ⇒ A1 A5 B-multi-b C5 red · audience `user` ⇒ A1 red · no boundary record of the walltime crossing ⇒ A4 red · cost in USD ⇒ A2 red · the strict test loosened to `<` ⇒ D1 red · the cap arm's budget read removed ⇒ B-multi-a red.
|
|
787
|
+
|
|
788
|
+
### A deployment whose runs declare no instruction family is told once, to each notice receiver, that no run reads an instruction file — `instructions.family_undeclared` (#4179; test c767c; behavior change: one operator notice, no verdict and no default reading moves; narrowing: none; widening: none; hard-breaking: one notice word, type level only (below); closed sets: `EngineNoticeCode` +1; exports: none added, none removed; wire: one notice code, additive) (@server @cli @client-core @test)
|
|
789
|
+
|
|
790
|
+
- **Before.** With the project-memory seats retired (7.67.0) and `TaskSpec.instructionFamily` not declared, a run completed with no instruction file in the model's first request and nothing anywhere said so: the wiring manifest's one word was `projectMemory: "absent"`, the notices of the run were whatever else the run had to say. A deployment that dropped its old seat and did not declare the family lost its CLAUDE.md silently (test c767c: `wiring_manifest = { projectMemory: "absent", instructions: null }`, one unrelated notice, `config.read_deny_floor_unkept`).
|
|
791
|
+
- **Now.** The project-context phase says it: one notice, `instructions.family_undeclared`, audience `operator`, minter `engine`, `detail: { sessionId }` of the run that first met the absence. The sentence says that no run reads any instruction file and names the key that makes them read one (`TaskSpec.instructionFamily`: `files`, with `rulesDir`, `alternate`, `includes`, `hostRules`). **Once per notice receiver** (a record keyed on the `onNotice` function the deployment wired, `noticeOncePerSink`): a later run, session, delegated leg — or further Runner built over the same function, a host that builds one per call included — says nothing more, a hot swap of the deps included; a Runner wired to another function says it once itself. With no `onNotice` wired the line reaches the default notice seat once per process, a console line like any notice without a sink. (Keying on the Runner was the first reading and is gone: a host that builds a Runner per call over one receiver heard it on every call.) A request that set `projectMemory: "disabled"` announces nothing (the request said it on purpose).
|
|
792
|
+
- **Not changed.** What a run without a declared family reads (nothing), the manifest's words (`"absent"` / `"on"` / `"disabled"`), the refusal of a retired seat by name (`config.deps_retired_key`), every run's verdict. A declared family, under either word of the off switch, adds no notice.
|
|
793
|
+
- **Upgrade order (INTEGRATION-CORE §7.1).** An engine before 7.67.0 refuses a spec that carries `instructionFamily` (`config.spec_unknown_key`, read by test on 7.66.1), so the declaration cannot go ahead of the upgrade: **upgrade the engine first and declare the family after, or in the same step.** 7.66.x does not learn the key.
|
|
794
|
+
- **Hard-breaking (type level only).** (1) *What changed:* `EngineNoticeCode` gains `instructions.family_undeclared` (`ENGINE_NOTICE_CODES` +1; the release's count, 135 → 153, is in the #3832 section). *Who is hurt:* a consumer that keys a record by `EngineNoticeCode` — a notice allowlist or router, an audience table, a renderer — gets a compile error until it gives the word a row. The wire is additive (`gate:wire` reads a closed set that gained a member as additive); a reader that closes on a default arm sees one more word. *Migration:* add the row (operator audience; render `message`; the session is in `detail.sessionId`).
|
|
795
|
+
- **Rules.** Announce-once units on the structured-notice lane 0 → 1 (`noticeOncePerSink`, the twin of the operator sink's record: one `WeakMap` keyed on the receiver function and one process-level set for the default seat, a test seam `__resetNoticedOncePerSink` beside it); readings of "no instruction family declared" 1 (the manifest's `absent` word and the notice read the one predicate, the project-context phase's `projectMemoryFact`). Added: one notice code, one dedupe verb. Retired: none.
|
|
796
|
+
- **Closed sets.** `EngineNoticeCode` +1 (three tables in `src/core/governance-codes.ts`: the code array, the audience table (`operator`), the minter table (`engine`); docs/CLOSED-SETS.md).
|
|
797
|
+
- **Consumers.** @server @cli @client-core: give `instructions.family_undeclared` a row in the notice table (operator audience; `message`; `detail.sessionId`); @server: when the deployment's file list maps to `instructionFamily`, the notice is the proof the mapping is live (it stops once a run declares), and the provision order is upgrade first. @test: c767c re-read — the same spec without `instructionFamily` carries exactly one such notice per receiver across two runs and across Runners built per call over the same receiver, and none once the family is declared.
|
|
798
|
+
- **Pins.** `test/backlog4179-instruction-family-undeclared.test.ts` — U1 two runs of one Runner, nothing declared ⇒ exactly one notice, operator audience, the first run's session, the key named, both runs completed · U2 declared ⇒ none · U3 first run declares, a later one does not ⇒ told once at the later run, a declared run after adds none · U4 `projectMemory: "disabled"` ⇒ none, a plain request on the same Runner ⇒ told · U5 a Runner wired to another receiver says it again, one wired to the same receiver does not · U5b one receiver, three Runners built in turn (a host that builds one per call), one run each ⇒ exactly one · U6 no `onNotice` ⇒ one console line across two runs and across two Runners (the process) · U7 a declared role map (each leg then reads its own rebinding of the deps) ⇒ still one · U8 a hot swap between two runs ⇒ still one. On the base tree (mutation M5, the mint removed) eight of nine are red (U2, the declared control, is green). Mutations, per fixture: the record removed ⇒ U1 U3 U5 U5b U6 U7 U8 red · the off-switch guard removed ⇒ U4 red · the condition flipped ⇒ all nine red · one record for every receiver ⇒ U5 red · the record keyed on the deps object ⇒ U5 U5b U6 U7 U8 red (a Runner per call, a declared role map and a hot swap each give a leg another deps object) · the default seat said every time ⇒ U6 red · the mint removed ⇒ U1 U3 U4 U5 U5b U6 U7 U8 red.
|
|
799
|
+
|
|
800
|
+
### The four store waits between a durable park and the published result run under the settle path's own bound — a parked run's result is no longer held by a checkpoint, session, memory-gate or cross-boundary-record call that never answers (#4126; behavior change: `CheckpointStore.setPendingSteer`, `SessionStore.touch`, the consolidation gate read and `crossBoundary.settle` on a run's way out are abandoned at the settle path's teardown bound where they used to be awaited without one; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @test)
|
|
801
|
+
|
|
802
|
+
- **Before.** A leg that parks on an approval commits its row, then walks the settle path to publish its result. Four calls on that walk were awaited with no time bound: the migration of the caller's undelivered steer onto the row (`setPendingSteer`), the session's activity touch (`SessionStore.touch`), the memory-consolidation gate read on the terminal-harvest tail, and the cross-boundary record's write at the leg's end (`crossBoundary.settle`: a session read, then an append). A store that never answered any one of them held the run's result (and the session lock) for ever: the park was durable, the caller never saw `paused`. Every other leg of that path (the date anchor, the harvest, the unpin, the environment teardown) already ran under the bound.
|
|
803
|
+
- **Now.** The four calls go through the one bound those legs have (`raceTeardownLeg`, `TEARDOWN_LEG_MAX_MS`, 15 s): a call still unanswered at the bound is abandoned (it keeps running detached, as the other legs' do) and the run publishes. Only that timeout is new: a call that throws or rejects inside the bound takes the path it always took (below). The sites that answer only the timeout with a report or a reading — the gate read, the cross-boundary record, and the parked leg's background-agent wait that already did — share one internal function, `raceTeardownLegOr` (the bound's own timeout goes to the caller's report; anything the leg itself throws comes out unchanged), instead of three copies of its try / catch. The touch, whose failure and timeout go to the same sink, is the bound with that sink as its catch. (1) `setPendingSteer`: the timeout is read as a store outage — reported to `onError` (phase `config`), the migration stops, and the steers it had not carried are announced through the existing `task.user_steer_undrained` account. (2) `touch`: reported to `onError` (phase `config`). (3) The gate read: read as "nothing owed now" — reported to `onError` (phase `config`), no consolidation run starts from this tail, and the recommendation stays on the gate row for the next leg's tail to find. (4) The cross-boundary record: read as unwritten and said through the record's own notice, `config.cross_boundary_record_unwritten` (`detail.phase: "settle"`, the bound's sentence in `detail.cause`; one per state); on a governed tree the next leg then reads the content after the newest record as off-target, as that notice already says for a failed settle write.
|
|
804
|
+
- **Not changed.** A store that answers or fails inside the bound behaves as before, failure included: the same reports, the same carried steers, the same record, and a failure the run's tail backstop used to receive (a `touch` that throws while the error sink itself throws; a gate read that rejects on the engine's face; a cross-boundary write that rejects) is still received by it — the bound turns its own timeout into a report and nothing else into one. No new notice code, no new knob, no new bound. A gate STORE that throws is still folded into "no row" inside the engine's own `readConsolidationGate` and stays silent, as before (the discriminating face is `readConsolidationGateState`). The final usage charge on the same walk is still awaited to completion (a different decision, its own ticket). `MaterializedMcp.dispose` and `MaterializedA2a.dispose` are read here, not changed: the host-pipe form's close is not waited for (a pipe whose close never settles leaves the result out at once, pinned as a reading), the SDK's stdio and HTTP transports bound their own close, and the A2A dispose sets a flag.
|
|
805
|
+
- **Who is hurt (narrowing, open to veto @server).** A deployment whose checkpoint store, session store, memory governance store or session-branch read takes longer than the bound on one of those four calls: the run used to wait through the slowness and now publishes without the call's effect — (1) a parked steer is announced undrained (sending it again is the caller's), (2) the session's activity clock is not refreshed this leg, (3) no consolidation run starts from this tail, (4) the record is missing. The abandoned call is not cancelled, so a `setPendingSteer` that lands after the result leaves the steer on the row as well as in the undrained notice.
|
|
806
|
+
- **Compensation.** None needed on a healthy store: 15 s is two orders above a row write. The undrained steer's notice names the inputs by `inputId` where the sender gave one, and the cross-boundary notice already carries the operator-side words for a missing record.
|
|
807
|
+
- **Rules.** Store waits between a park and its published result with no time bound: 4 → 0 (the four join the one bound; no second bound, no new code). Special cases: 0.
|
|
808
|
+
- **Pins.** `test/backlog4110-park-to-publish-store-waits.test.ts`: cells 2 3 4 5 red on the base (the result is not out inside the 1.5 s window while a 200 ms budget is in force) and green on the branch, each reading the bound's own report (the typed timeout on `onError`, or the record's notice); 4 also pins that the abandoned read's late answer starts no consolidation run. Cells 0 0b 1 4b 6 are controls and readings, green on both: 0 nothing held; 0b a site that is already bounded (the date anchor's leg) read through the same frames; 1 the final usage charge, still unbounded by design; 4b the gate answering at once with an owed recommendation does start the run (the witness that 4's "no run" reads the abandoned answer, not an empty gate); 3b a `touch` that throws ⇒ reported once with the store's own error, result out at once; 3c a `touch` that throws while the error sink itself throws ⇒ the run ends `failed` through the tail backstop (a bound that swallowed the throw would publish `paused`); 4c the engine's gate face rejecting ⇒ not read as "nothing owed", the rejection travels (the run ends `failed`, no run starts); 4d the gate store throwing under the engine's face ⇒ folded, nothing reported, result out, no run; 5b the cross-boundary read throwing ⇒ the lane's own one notice, result out at once; 6 `MaterializedMcp.dispose` over a host pipe whose close never settles. `test/backlog4126-race-teardown-leg-or.test.ts` pins the one function the gate read, the cross-boundary record and the parked leg's wait share (`raceTeardownLegOr`): the timeout goes to the caller's report or reading; a leg that rejects, throws before it returns a promise or throws another bound's timeout comes out unchanged; a report that throws is not hidden. Mutations: fifteen, each red only in the cells of the site or the function it touches (the touch's swallowing form also turns `task-end-backstop.test.ts` F-05 and W8 red; the function swallowing every failure turns three direct cells and 4c red).
|
|
809
|
+
- **Consumers.** @server: a checkpoint or session store that stops answering on those four calls no longer holds a parked run's result or its session lock; the run publishes `paused` at the bound with the incident on `onError` (or the record's notice). @test: black-box form — park a leg on an approval, make one of the four store calls never answer, read the result frame at about 15 s.
|
|
810
|
+
|
|
811
|
+
### A session's file history lives at its address — `fileHistoryScopeOf(sessionId, occupant)` names the history scope of every row, the address of (id, occupant) for a row born in the address epoch and the bare id for every other row; new occupant tokens carry the epoch mark `.2`; the fork carry and the deletion cascade read the same function (#3996, #3992 S1; ruled C-R1335 and its first note; behavior change: an occupant created under a reused id no longer reads, rewinds to or rewrites the history its predecessor recorded; widening: none against 7.67.1 — R-1 / R-3 below keep two 7.67.1 readings, open to veto @server; narrowing: three (N-1 / N-2 open to veto @server; N-3 an id that ends like an address on a row outside the epoch is refused file history, open to veto @server @cli @client-core); hard-breaking: two (HB-2 the history scope is no longer always the session id; HB-3 `SessionStoreSummary` gains two required members), each open to veto @server @cli @client-core @sdk; token shape change (TF-1, within the contract), open to veto @server; closed sets: none; exports: +`fileHistoryScopeOf`; wire: none — the token never goes on the wire) (@server @cli @client-core @sdk @test)
|
|
812
|
+
|
|
813
|
+
- **Rule.** One function, `fileHistoryScopeOf(sessionId, occupant)` (root export), is the file-history scope of a session row, where `occupant` is the row's `SessionMetadata.incarnation`. A token the bundled mint point minted (`incarnationCell`: a lowercase v4 followed by `.2`, 38 characters, matched whole) says the row was born from that write, so its scope is `sessionId~` + the first 22 base64url characters of the token's SHA-256 — two occupants of one id never share a scope, and the token is not recoverable from it. Every other occupant — none, a token minted before the mark, a host's own token — reads as the bare id, where that row's history has always been. Such a row whose id itself ends like an address (`~` + 22 base64url characters) has no scope — its bare id could be another row's address — so the function refuses it (`config.invalid_argument`, N-3); an id with `!`s after that tail is refused the same (one rule for the shape). So no scope is ever two rows'. The mark is spelled and read in `repo-utils.ts` only.
|
|
814
|
+
- **Before.** File history was keyed by the session id alone. An id erased and created again with the history not erased (the built-in `TtlSessionStore`'s default `evict: "delete"` drops the row and not the history; a host store with no `deleteBySession` leg) left the predecessor's graph under the id: the newcomer rewinding to its own turn could rewrite a file only the predecessor had edited, and a fork's lazy carry could take another occupant's graph.
|
|
815
|
+
- **Now.** Every engine seat names history by its row's address: the run's first-touch record and its retraction, the turn boundary, the restore (`rewindFilesTo`, `resumeAt` + `restoreFiles`, `Runner.rewindFiles` and its preview), the same-tree child (it records into the root's address), and the fork carry. The run's own scope and its session id are separate: refusals, `rewind.boundary_unrecorded`, the read-state channel and `onError` contexts still name the session id. The carry reads the source address from the fork row's lineage cell, `fileHistoryScopeOf(forkedFrom, forkedFromIncarnation)`, into the fork's own scope, and reads no row of the source; a fork row recording no source occupant carries nothing and says so once per (history store, fork scope). The carry memo is keyed by the fork's scope, so a fork id created again carries for itself. `RETENTION_CASCADE` steps 9 and 10 read the same function (below).
|
|
816
|
+
- **Hard-breaking HB-2 (open to veto @server @cli @client-core @sdk): the history scope of a session is no longer always its session id.** Now: every host call that passes a session id where a file-history scope is meant passes `fileHistoryScopeOf(sessionId, the row's incarnation)` instead — the token taken from the same lock section, commit or head read as the row, and the same cell the engine acquires; for a row born before the epoch it computes to the id, for a new row to the address; only the file-history argument changes, the policy, anchor and box legs of the same call keep the id. Who is harmed: the server's five host seats (seven coordinates) — the local erase leg `session-erase.ts:132-145` (including the `reap(sessionId, [])` arm), the SQL retention purge `retention-store-sql.ts:855-889` (the sentinel lock and the six-table DELETE take the scope set; the pending rows already carry `{ id, incarnation }`), the sync export `session-sync.ts:75-77` (callers: the manifest export and the blob pull boundary `http/routes/session-sync.ts:408`), the sync replay `session-sync.ts:151-177` (caller `routes:324`; the target scope is the committed row's token), and the overwrite wipe `session-sync.ts:424-435` `wipeSessionAttendants` (caller `routes:882`; the wipe uses the token before the replacement, the replay the one after). Left unchanged, only sessions created after the upgrade are hit: deleting one leaves its history and raw bytes, export answers no history, blob pulls 404, an overwrite leaves the old graph. Compensation: change those seven coordinates to call the function.
|
|
817
|
+
- **Hard-breaking HB-3 (open to veto @server @cli @client-core @sdk): `SessionStoreSummary` gains two required members, `incarnation: string | null` and `forkedFromIncarnation: string | null`.** Now: each listed row carries its occupant cell and the occupant of `forkedFrom` it was cut from (`null` when the row has none), read in the same row read as `forkedFrom`; the deletion cascade computes each fork's scope from them. Who is harmed: a host that implements `SessionStore.list` itself (a compile error in TypeScript; a JavaScript host's rows without the keys make the cascade coordinator refuse before any history is touched — the contract kit's `occupantKeys` case). The server implements no `list` of its own (its fork-routing store passes the bundled `TtlSessionStore` projection through), so it has no code to change. Compensation: project the two cells as `TtlSessionStore.list` does; they are store-interface material and never go on the wire.
|
|
818
|
+
- **Token shape (TF-1, within the contract, open to veto @server).** Now: `incarnationCell` mints `<v4>.2` (38 characters) for a new row instead of a bare v4 (36); every existing row's token is unchanged, and the contract still says only "a non-empty string, minted once, compared, never on the wire". Who is harmed: a host that stores tokens in a uuid-typed or 36-wide column or validates them as canonical uuids (none in the server's sources); four server comments that say the token is a v4 and any test pinning the whole token shape; a host that gives a token to a row that already holds engine state by calling the birth mint (the server's local store when it fills the first cell of an archive with no token) moves that row into the address epoch, and the history it recorded at the bare id is no longer read. Compensation: store tokens as opaque strings of at most 64 characters; give such a row its earlier token, or a token of your own without the mark — never `incarnationCell(undefined)`, which is for the write that creates a row.
|
|
819
|
+
- **R-1 — kept as in 7.67.1 (open to veto @server).** A fork whose source row is gone (deleted with its history kept, or a session store with no `headOf`) carries the source address it records: no row is read. 7.67.1 carried such a fork too (by its id); the source-row comparison #3982 first made, which would have refused it, did not ship (that section, above).
|
|
820
|
+
- **R-3 — kept as in 7.67.1 for a source from before the epoch (open to veto @server).** A fork that recorded a source occupant from before the epoch carries the source's bare-id scope — the scope 7.67.1 carried every fork from — without reading who holds the source id now (no fork reads it): a source id erased and created again before the upgrade, whose newcomer wrote history, hands that mixed graph to a fork whose first turn comes after the upgrade (`docs/KNOWN-LIMITS.md` KL-3992-a). Bounded for the bundled stores (no row is born before the epoch any more); unbounded on a host minting its own tokens.
|
|
821
|
+
- **Narrowing N-1 (open to veto @server).** A run whose own row cannot be read (`getMetadata` throws) has no history scope: its first-touch backups answer `track_failed` (refused under the default `onTrackFailure`, passed and told under `"proceed-unprotected"`), it records no boundary, and a rewind is refused `rewind.restore_failed` (the run seat and `Runner.rewindFiles`). It used to record at the bare id, which belongs to another occupant when its row was born at an address.
|
|
822
|
+
- **Narrowing N-2 (open to veto @server).** The cascade's step 10 carries a session's history only into the forks whose row records that session's occupant as their source (the source address equals the scope step 11 deletes); a fork row recording no source occupant is no longer carried into when its source is deleted (the engine's own carry never carried it either).
|
|
823
|
+
- **Narrowing N-3 (open to veto @server @cli @client-core): a session whose id ends like an address and whose row was not born in the address epoch has no file history.** Now: for such a row (`~` + 22 base64url characters at the end of the id, or that followed by `!`s; its token absent, from before 7.68.0, or a host's own) `fileHistoryScopeOf` throws `config.invalid_argument` ("a caller-chosen id collides with the address epoch; re-id the session") and nothing of any history is read, written or deleted for it — its first-touch backups answer `track_failed` (refused by default), it records no boundary, `Runner.rewindFiles` and a run's `rewindFilesTo` fail with that code, a fork of it carries nothing (one `adopt_failed`), the cascade's history legs select nothing for it, and the history it recorded before the upgrade stays where it is (KL-3992-a (3)). Who is harmed: only a host that lets callers choose session ids and whose session repo accepts `~` in one (the bundled `FileSessionRepo` takes only `[A-Za-z0-9_.-]`, so no such row exists there; the in-memory repo and the example SQL adapters accept it) — @server: the sessions its stores keep from its releases up to 7.125.x, whose submit door took a caller-chosen session id of up to 64 characters of any shape (its 7.126.0 made the door take canonical uuids only and kept the existing rows): such a row whose id ends in `~` and 22 characters of `[A-Za-z0-9_-]` (then any `!`s) was born before the address epoch, so its erase, retention-purge and sync legs, once they call the function (HB-2's seven coordinates), meet `config.invalid_argument` for it, and its history at the bare id reads through `readFileHistoryScope` as an address of the shorter id (KL-3996-a); the uuids its door has taken since are not hit; the engine mints uuidv7; @cli and @client-core where a user-chosen session name or id is passed through as the session id (inferred, not read in their sources); an id shaped this way loses rewind until it changes. What to do: give the session a new id (its new row is born in the address epoch); a host that erases or syncs history catches the refusal and skips the history leg for that row. Rejected alternative: renaming the bare scope (one more `!`) — it let a pre-epoch `X~<22>` read, write and delete the history of a pre-epoch `X~<22>!` (C14c's baseline).
|
|
824
|
+
- **Mixed versions.** A session created by this build and served by an earlier one (a rolling upgrade, a rollback) has its history written at the bare id by the earlier build and read at the address by this one: each side rewinds only to the boundaries it wrote ("no boundary", never a wrong restore).
|
|
825
|
+
- **Not changed.** History stores, their verbs, their keys and their schemas (the scope stays an opaque string; every bundled store and the server's SQL stores keep their code); every existing row's token and scope (zero migration) but N-3's; the refusal codes (N-3 reuses `config.invalid_argument` on the history seats), and their sentences but one — `rewind.child_scope_unsupported` now says "its ROOT session's file history (scope "…")" instead of calling the scope a session id; the own-log check of a rewind target (#3994).
|
|
826
|
+
- **Rules.** Scope readings: one (`fileHistoryScopeOf`). Readers of the mark: one. The carry's refusal arms: one (no recorded source occupant) — against 7.67.1's none and the stage's three (same occupant / other occupant / row gone, plus a missing-`headOf` arm, #3982's first form, which did not ship); its `sessions` parameter is gone. The cascade selects forks and deletes history with the same function. No old token is read two ways: its scope is the id, a new token's is its address, each row has exactly one.
|
|
827
|
+
- **Pins.** `test/backlog3992-file-history-occupant-address.test.ts` C1–C19 (the reused id on the verb and the run seat, partitions, the fork carry's shapes, rows from before the epoch, the phase with an unreadable row, the function's table and the marker's privacy, the A / B collision shape refused on every seat (C14c), the pre-epoch fork and cascade regressions); `test/backlog3005-session-deletion-cascade.test.ts` (the kit judges history by scope: cases `scopes` and `occupantKeys`; defects `erase-by-id`, `source-unchecked`, `q1-ignored`, `keys-defaulted`); re-pinned to the address: `backlog3982-*`, `backlog3994-*`, `backlog535-*`, `backlog1048-*`, `backlog1050-*`, `backlog1130-*`, `backlog1298-*`, `rewind-files`, `backlog1468-k2-laws` (the minted token's shape), `backlog1468-s1-session-store-kit`.
|
|
828
|
+
- **Divergence from the reference (D-3992-a, ruled C-R1335, open to veto @server).** The reference keys a session's file history by the session alone; this engine keys it by the session's address, (id, occupant), for a row born in the address epoch, so an occupant created under a reused id never reads, rewinds to or rewrites its predecessor's history (N-3 is the one narrowing that comes with it).
|
|
829
|
+
- **Consumers.** @server: HB-2's seven coordinates (S-3774); TF-1's four comments, the token-shape test and the local store's first-cell fill (R18) and overwrite reuse (R19: wipe with the token before the replacement, replay with the one after); R-1 / R-3 / N-1 / N-2 / N-3 above (N-3: its stock of caller-chosen ids from releases up to 7.125.x — census it with `sessionIdHasScopeTail` and re-id the rows it finds; ids its door has taken since 7.126.0 are uuids). @cli / @client-core: N-3, where a session id is chosen by the user. @sdk: `SessionStoreSummary`'s two members in the declaration files. @test (black box): B1 a session from before the upgrade rewinds to a turn before the upgrade; B2 an id reused after the upgrade (history kept) — the newcomer's rewind leaves the predecessor's file alone; B3 a fork carried, its source deleted, a new worker process — no `adopt_failed` line, its pre-fork boundary restores; B4 a fork carried before the upgrade, its source edits a file first after it — the fork's rewind leaves that file alone. @cli @client-core: no read of either cell.
|
|
830
|
+
|
|
831
|
+
### A rewind goes only to an entry of the session's own log — `Runner.rewindFiles` `to` and `TaskSpec.rewindFilesTo` are refused `rewind_snapshot.unresolvable` when the session's log does not hold the entry, before the history store is asked (#3994; ruled C-R1335; behavior change: a rewind to a boundary only another occupant of the session id recorded no longer moves files; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @test)
|
|
832
|
+
|
|
833
|
+
- **Rule.** A boundary is keyed by an entry of the session's own log, so a target the log does not hold is not a boundary of this session. One reading — `session.getEntry(target)`, the reading a branch target (`resumeAt`) is already read by — gates every seat that rewinds files before the history store is asked: the out-of-run verb (its preview and its restore) and the run's `rewindFilesTo`. The seats cannot disagree about whose boundary a target is.
|
|
834
|
+
- **Before (7.67.1).** The file-history store is keyed by the session id alone, and both seats asked it about whatever entry id they were handed. An id erased and created again with the history store not told (the built-in `TtlSessionStore`'s default `evict: "delete"` drops the row and not the history; a host whose store has no `deleteBySession` leg) left the earlier occupant's boundaries under the id, and the new occupant could name one and have files it never touched rewritten.
|
|
835
|
+
- **Now.** A target the session's own log does not hold is refused `rewind_snapshot.unresolvable`, in the sentence an entry with no boundary gets (the preview in its own sentence), so no wording tells the caller that someone else's boundary is there. No history verb is asked about the target and no file is moved. The verb's check runs after the request, deployment, session and no-env doors (a missing session is still `resume.session_not_found`, an unwired store `rewind.store_unconfigured`, no filesystem env `rewind.files_env_unsupported`) and before the verb's fork carry; a run's own fork carry (made on every run of a fork, whatever it rewinds to) still comes first.
|
|
836
|
+
- **Narrowing (open to veto @server).** A target that is not in the session's log refuses where it used to rewind: a boundary only an earlier occupant of the id recorded, and a fork's carried copy of a source boundary recorded after the cut (the fork's log never held that entry). A fork's own copied log holds every boundary entry cut before it, and rewinds to them as before; a backend that wakes a bounded window must answer a below-floor id through `SessionStorage.getEntry` (now written on the member; `resumeAt` already read it so), or a rewind to a turn below its floor refuses where it used to rewind — the SQL windowed stores in the server package do (`readBelowFloor`), a third-party windowed storage that answers only its loaded window does not.
|
|
837
|
+
- **Not changed.** The history store, its keys and its verbs; the refusal codes and their sentences; an own-boundary rewind on a plain session, on a re-created id and on a fork; the history scope (this section does not change it; #3996, above, makes it the row's address, so a re-created id born in the address epoch no longer shares its predecessor's history — a row from before the epoch keeps the session id as its scope, KL-3992-a).
|
|
838
|
+
- **Rules.** Added: one function (`refuseTargetOutsideOwnLog`) called at the two seats. Retired: none. No new code, word or member.
|
|
839
|
+
- **Pins.** `test/backlog3994-rewind-own-log.test.ts` — G0 the premise (the store holds the earlier occupant's boundary under the id; the new occupant's log does not hold its entry) · G1 / G1b / G1c the new occupant refused the earlier occupant's boundary on the verb, its preview and the run's `rewindFilesTo`, no file moved · G2 controls (an own boundary on a plain session through the verb, the preview and the run seat; on a re-created id) · G3a controls, answers unchanged (an entry with no boundary and an absent entry say one sentence; `rewind.invalid_spec`, `resume.session_not_found`, `rewind.store_unconfigured` as before) · G3b a foreign entry is refused in that same sentence, restore and preview alike · G4 control: a fork rewinds to a boundary entry its copied log holds · G4b a fork is refused a boundary the source recorded after the cut.
|
|
840
|
+
- **Consumers.** @server (the rewind read door): a `to` the session's log does not hold answers `rewind_snapshot.unresolvable`, the code a missing boundary already answers — nothing new to map; a third-party history store with no `deleteBySession` leg no longer lets a re-created id rewind to its predecessor's boundaries; a windowed session storage must answer `getEntry` below its floor (the SQL stores do; `LocalSessionWindow` is a read-only view that answers its window only — a run appends, so it is not what a run acquires, inferred from its refusing `appendEntry`, not run). @test (black box G1): delete a session's row keeping its history, create the id again, have the new session name the old session's turn entry — refused, file bytes unchanged. @cli @client-core: no read.
|
|
841
|
+
|
|
842
|
+
### An allow rule auto mode does not use is said to the person once per conversation, and the rule read faces say it too: `permission.allow_rule_not_used_in_auto`, `EffectivePermissionRule.autoStrip`, `allowEntryAutoStrip` (#1367; user-visible disclosure; hard-breaking: one notice word; says more than the reference — open to veto @cli @server @client-core @test)
|
|
843
|
+
|
|
844
|
+
- **Rule.** Auto mode stands down the allow entries that would let a call skip the classifier — the whole shell, a program that runs whatever code it is given, every shell entry under `TaskSpec.classifyAllShell` — one judgment (`src/core/auto-strip.ts`, moved out of `allow-entry.ts` with no change to what it strips), which now names why: `AUTO_STRIP_REASONS` = `whole_shell` / `code_runner` / `classify_all_shell`. Every reader of the fact takes that one judgment: the strip itself, the person's notice, a stored row's read face, a list entry's read face.
|
|
845
|
+
- **Before.** A person whose saved `Bash(npm run:*)` (or `Bash`, `Bash(python:*)` on the request's list, a whole-tool rule naming a tool that answers to the shell under another name) is stood down in auto mode saw every matching command go to the classifier with no word anywhere: test emb1 (cli 1.0.165 / core 7.65.0) read zero notices on the TUI transcript, the `-p` frames and `engine_notice`, for both the settings form (the request's `allowedTools`) and the card's "don't ask again" form (a stored row); the code set had no such word; the rule read faces said nothing a shell could mark.
|
|
846
|
+
- **Now.** (1) The first call the gate judges in `auto` that a stripped ALLOW entry would have been read for — a stored row's (one whose scope applies on the leg and whose text reads as an allow rule), the request's or the deployment's `allowedTools` entry's, an active skill's `allowed-tools` entry's — tells the person `permission.allow_rule_not_used_in_auto` (audience `user`, minter `engine`), once per conversation for that rule (the session's announced record — `tellOncePerSession`, the one run-time face of the once-per-session record, which `model.capability_downgraded` and `inherit_env.names_withheld` now ride too). `detail: { entry, source, reason, toolName, toolCallId, skill?, sessionId, sourceSessionId?, agentId?, runId }`; `source` ∈ `user` / `project` / `session` (the stored row's `EffectivePermissionRule.source`) ∪ `request` / `deployment` (`AllowEntrySource`) ∪ `skill`. The sentence: `The allow rule "Bash(npm run:*)" (saved in your user rules) is not used in auto mode: it allows a program that runs whatever code it is given, which is too broad for auto mode to honor, so in auto mode it approves no command on its own: unless another rule approves one, it goes to the auto-mode classifier, which may block it. To have a command approved without the classifier, save a narrower rule that names the whole command, such as Bash(npm run build); or switch to the default permission mode, where this rule applies as written. Said once per conversation for this rule.` (2) `EffectivePermissionRule.autoStrip?: "whole_shell" | "code_runner"` on the store's `effective()` rows, derived beside `source` and `status`. (3) `allowEntryAutoStrip(entry)` on the package root reads a list entry (a settings `permissions.allow` string) the same way. Both read faces judge without a run: no call, no leg declaration.
|
|
847
|
+
- **Says more than the reference (D-1367-a, open to veto).** The reference strips at mode entry and writes only a debug line (`Ignoring dangerous permission … (bypasses classifier)`, cli321.js:292566); its user-facing word is the `/auto-mode-setup` wizard's "Review rules that skip checks" page (cli321.js:312058, opened at cli321.js:317504). sema has no such wizard and its TUI defaults to auto, so the run says it.
|
|
848
|
+
- **Not changed.** What the strip stands down (every reading of `autoStripped` / `strippedInAuto` is the same); `ruleOffers` in `default` still offers the tier auto mode would strip (the reference composes such a Bash rule outside auto mode too, cli321.js:307154); default mode tells nothing; `compileAllowList` rows carry no new member (they are also the session narrowing bucket's reading, which is never stripped).
|
|
849
|
+
- **Hard-breaking.** *What changed:* `EngineNoticeCode` gains `permission.allow_rule_not_used_in_auto` (`ENGINE_NOTICE_CODES` +1; the release's count, 135 → 153, is in the #3832 section). *Who is hurt:* a consumer that keys a record by `EngineNoticeCode` — a notice allowlist or router, an audience table, a renderer — gets a compile error until it gives the word a row. *Migration:* add the row (user audience; render `message`, or the rule, where it is kept and why from `detail`).
|
|
850
|
+
- **Rules.** Strip judgments 1 → 1 (moved, named); its readers that answer "is this rule used in auto" 1 → 3, all calling the one judgment, pinned equal over one batch (F3). Run-time once-per-session copies 3 → 1 (`tellOncePerSession`). Added: one notice word, one closed set (`AutoStripReason`), one read-face member, one root function.
|
|
851
|
+
- **Closed sets.** `EngineNoticeCode` +1 (the code array, the audience table, the minter table); `AutoStripReason` new (`AUTO_STRIP_REASONS`, three words; the read faces carry the first two).
|
|
852
|
+
- **Consumers.** @cli (L-1069 / L-2095 / L-2097): render the notice; on `/permissions` mark a saved row by `autoStrip` (from `GET /v1/rules`) and a settings `permissions.allow` entry by `allowEntryAutoStrip` — the shell judges nothing itself; the store refuses `Bash(python:*)`-style interpreter prefixes as an allow, so that shape only reaches the engine on a list; a stored allow row the grammar reads as no rule grants nothing in any mode and is not named. @server: map `autoStrip` into the `GET /v1/rules` row beside `source` / `status` (and leave it out of the import face, a derived key); give the notice code a row. @client-core: the notice code's row; `allowEntryAutoStrip` beside the `compileAllowList` port. @test: re-read emb1 ① (the notice on both forms) and ③ (the read face).
|
|
853
|
+
- **Pins.** `test/backlog1367-auto-strip-notice.test.ts` — N1 auto × stored wide + narrow: one notice for the wide rule on its first call, none for the narrow rule, none again in the run · N1b none in a second run of the conversation · N2 default: none · N3 the four shapes (`Bash(npm run:*)`, `Bash`, `Bash(python:*)` on the request's list, `sh` aliasing the shell) · N4 the request's list (emb1 s1a) · N5 narrow only: none · N6 `classify_all_shell`: the mode as the only way out · F1 store rows · F2 the export · F3 the three readers alike · C1 the code's place.
|
|
854
|
+
|
|
855
|
+
### A spawn name that already belongs to an agent is not taken over — the second `Agent({ name: "scribe" })` is `scribe-2`, its launch card says `name: scribe-2`, and `scribe` still reaches the first (#1696; behavior change: model-visible card line + `details.name` added; hard-breaking: none; exports: `TaskRegistry.claimAgentName` added; wire: `details.name` on the `async_launched` result)
|
|
856
|
+
|
|
857
|
+
- **Rule.** A name binds ONE agent in the pool a caller resolves names in. The registry binds spawn names (`TaskRegistry.claimAgentName`): the pool is the delegation tree's — every background-agent row the tree's root session can see whose handle this process holds, any status (a finished agent still answers to its name while its handle is held; a durable row whose handle this process does not hold — after a restart, or once the handle is evicted — is not in the pool, and a new spawn can take its name), so a child spawning a name the root's or a sibling's agent answers to does not rebind it — plus the names launches from that tree hold before their rows exist. A requested name already in the pool takes the first free `<name>-N` (N ≥ 2; case-blind, as names resolve), keeping the requested spelling and cut so the result stays within the 64-character spawn-name grammar (a revival and the durable box re-read the row's name through it). The claim is held per delegation call and released when that call ends, so two spawns of one name in one assistant message (the delegation tool runs in parallel) bind two names. A revival keeps its row's name.
|
|
858
|
+
- **Before.** The second spawn registered under the same name; its card was the same as the first's; name resolution (latest wins) then sent `SendMessage({ to: "scribe" })` to the newer agent with no word to anyone (test [10335] E1, bbq1 e1 on 1.0.165).
|
|
859
|
+
- **Reference.** CC 2.1.321 teammate spawn (`cli321.js` L310531: case-insensitive over the roster, `<name>-2`, `-3`, …; the spawn card prints `name: …`, L292500); the same form as 2.1.296. Not followed: CC's plain-subagent path with agent teams off overwrites the name (`registerName`, L292499) — sema's named spawn is its teammate lane (it always runs in the background), so the teammate form is the one that applies.
|
|
860
|
+
- **Card.** `asyncLaunchedReceipt` carries a `name: <bound name>` line under `task_id` for every named launch (both background lanes), as CC's card does; the result's `details.name` carries the same value. The card does not say why a name differs from the requested one (CC's does not either).
|
|
861
|
+
- **Who is hurt (open to veto @cli @client-core).** A host or test that spawned the same name twice and expected the name to move to the newest; a footer that labels a row by the requested name instead of the row's name. Compensation: the bound name is on the card, in `details.name`, on the registry row and every frame (`spawnName`).
|
|
862
|
+
- **Rules.** +1 (the binding rule, in one place: `claimBackgroundAgentNameLane`); the resolver is unchanged (latest wins stays the tiebreak for pools a binding cannot reach, e.g. rows registered before this version or a session-scoped and a task-scoped row of one session, RB-115).
|
|
863
|
+
- **Pins.** `test/backlog1696-spawn-name-binds-once.test.ts` — 1–5 and 8–11 red on the base (`34cd602b6`); 6 (another session's pool) and 7 (a refused spawn binds nothing) are controls, green on both. Mutations M1–M10 + M5b each turn a cell red.
|
|
864
|
+
|
|
865
|
+
### A stopped agent is listed with whether SendMessage can still continue it and for how long — `continuable for 29m` / `continuable (durable record)` / `not continuable` (#1699 core half; behavior change: model-visible ListAgents column and tool description sentence; hard-breaking: none; exports: `AgentContinuation` type, `ListAgentsRow.continuation`, `SubagentRetainLedger.continuableUntil`, `RetainLedgerSeats`, `liveRetainEntryOf`, `resumabilityClaim` moved into `retain-ledger.ts`; wire: `continuation` on ListAgents `details.rows[]`)
|
|
866
|
+
|
|
867
|
+
- **Rule.** The retain window is one fact on the ledger: `continuableUntil(entry)` = the entry's last settle + its retain window. Both reapers (the lazy sweep, the per-entry timer) evict past it. ListAgents answers each stopped agent (`finished` / `failed` / `killed`) through the ladder SendMessage walks: `liveRetainEntryOf` (session, run, sibling ledger; the first ledger that KNOWS the spawn decides, an eviction's tombstone included — the read SendMessage now shares) and `resumabilityClaim` (moved beside the ledger; the terminal frames read it too): a live retain entry ⇒ `{ by: "retained", until }` while the window has not run out; no entry, a row the durable rung serves (any row but a fork's, named or not — tier-3 revives by handle) with a confirmed durable record and a SendMessage tier that revives ⇒ `{ by: "durable" }`; otherwise (a user's stop, a spent resume budget, nothing retained or recorded) ⇒ `{ by: "none" }`.
|
|
868
|
+
- **SendMessage (edge, behavior change).** With an eviction's tombstone in an earlier ledger and a live entry for the same spawn id in a later one, SendMessage now tries the durable rung first (the live-entry read used to skip the tombstone while the ledger choice stopped at it, so it answered `resume.evicted` without trying tier-3). Spawn ids collide across ledgers only when two conversations reuse one tool-call id.
|
|
869
|
+
- **Open item (pre-existing, not changed here).** The terminal frames' `resumable` still answers for NAMED durable rows only (`durableRungServes: <named>`, pinned by `agent-resume-honesty` "an ANONYMOUS durable child reports false"), while tier-3 revives an unnamed row by its handle — the frame under-claims; the listing states the rung. Left for a ruling.
|
|
870
|
+
- **Before.** ListAgents listed a finished agent as `finished` and attached one tier-wide sentence to the count; the window existed only inside the two reapers' expiry predicates, and nothing told the model or a host whether THIS agent could still be woken (test [10335] E2, bbq1 e2: the name woke a finished agent at 280 s).
|
|
871
|
+
- **Reference.** CC 2.1.319 lists ended subagents under "Earlier in this session" as reopenable (changelog-321.tsv) and keeps an ended teammate `[idle]`; CC sets no retain window of its own. sema's default window (`RETAIN_DEFAULT_TTL_MS`, 30 min) is above the ≥ 6 min observed in CC; a durable record has no engine deadline.
|
|
872
|
+
- **Not changed.** Which agents are listed (live registry rows and roster rows, as before); the tier sentence on the finished count; SendMessage's ladder (it reads the same entry through the shared function).
|
|
873
|
+
- **Model-facing text.** ListAgents' tool description changes by one clause, on its background-agents sentence: `… the card id, state and deadline.` → `… the card id, state and deadline; for one that stopped, whether SendMessage can still continue it and for how long.` The prompt-assembly golden fixture (`test/prompt-golden-1302.json`, the hands-tools wire face) is refrozen for exactly that sentence.
|
|
874
|
+
- **Consumers (@cli @client-core).** The shell half of #1699 (L-1334) can render the footer from `details.rows[].continuation` (`until` is an epoch-ms instant). A host poll face (`TaskRegistry.list` rows) does not carry the fact — the ledgers are the delegation surface's, not the registry's.
|
|
875
|
+
- **Rules.** Expiry predicates over the window: 2 → 1 (`continuableUntil`). Reads of "the live retain entry of a spawn": SendMessage's inline chain → `liveRetainEntryOf` (1 reader → 2, one rule). Resumability: 1 predicate, readers 2 → 3.
|
|
876
|
+
- **Not breaking — the ListAgents mount's registry option (a wire-manifest reading).** `ListAgentsToolOptions.registry.getAccessibleTask` answers `{ name?, status, agentType? }` and `registry.list` rows carry `toolUseId?` (the listing's durable rung reads them: any row but a fork's); a host registry written against the earlier answer compiles and is read as before, and a missing `agentType` reads as "not a fork".
|
|
877
|
+
- **Pins.** `test/backlog1699-listagents-continuation-window.test.ts` — 1 and 2 on the real mount path (a Runner with a shell environment, the model calls ListAgents), 3–5 and 7–9 over hand-built rows and the ledger; 1–5 and 7–9 red on the base (`34cd602b6`), 6 (a running agent carries no continuation) green on both. Mutations N1–N11 each turn a cell red.
|
|
878
|
+
|
|
879
|
+
### SendMessage's success receipts name the agent and the delivery state, never its internal task id — `Message sent — agent "scribe" resumed in the background`, not `agent "scribe" (a3051d57849717e30) resumed …` (#4080; behavior change: the text of seven receipts drops the id (five success receipts and the two `Message not sent …` results that leave the message queued); narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@cli @client-core @test)
|
|
880
|
+
|
|
881
|
+
- **Before.** The receipts a delivery ends in — queued for delivery to a running agent, parked for a finished one, accepted-but-unconfirmed, revived from its durable row, resumed from its transcript, and the two results that say `Message not sent …` while the message stays queued for a later revival (the mailbox lease is held by another revival; the revival could not start) and carry no `isError` — spelled the agent as `agent "<name>" (<task id>)`, or `agent <task id>` when it was addressed by id. The Agent tool's own receipt marks that id "internal ID - do not mention to user"; the SendMessage receipt printed it into text the user reads (cli card; sema-test bbq1 on 1.0.165).
|
|
882
|
+
- **Now.** One function (`receiptAgentLabel`, `src/agents/send-message-tool.ts`) spells the agent for those seven: the name the caller addressed it by; when the caller addressed it by id, the row's chosen name, else its description; else `the agent`. The delivery state words are unchanged (`queued for delivery … at its next turn`, `parked`, `accepted … UNCONFIRMED`, `was revived in the background`, `resumed in the background`). The id stays where a host reads structure on the five delivery receipts (`details.task_id`, as before); the two `Message not sent …` results carry `details: { error, to }` and no `task_id` — as in 7.67.1, where the id was in their text, which no longer prints it (KL-4080-a, #4273).
|
|
883
|
+
- **Not changed.** The refusal sentences (every result with `isError`, e.g. `Message not sent: agent "<name>" (<task id>) is parked …`) keep the id: there it names the row that refused, and the model acts on it (which of two agents of one name answered). The correlation marker `[resume-…]` of the resumed receipt is a separate token the resumed child's reply carries; it is not a task id. The peer-session and `main` receipts name a session box, not an agent, and are not touched.
|
|
884
|
+
- **Who is hurt (open to veto @cli @client-core).** A host that cuts the task id out of the receipt's text loses it there; it reads `details.task_id` (no new key). No consumer in this repository did; two pins asserted the old spelling and now assert the new one.
|
|
885
|
+
- **Rules.** 2 → 1: the receipts spelled the agent through two label forms (`who`, `whoT3`) that both carried the id; the seven now read one function; the refusals still read those two. Special cases: 0.
|
|
886
|
+
- **Pins.** `test/send-message-tool.test.ts` (resumed arm by name and by id, running arm by name and by id, the structured result keeps the id), `test/rb423-sendmessage-parity-completion.test.ts` (revived arm by name and by id), `test/agent-revive-tier3.test.ts` (the revival that cannot start, and the lease held by another revival, each leaving the message queued — by name and by id), `test/backlog1697-named-agent-background.test.ts` and `test/sendmessage-agent-name.test.ts` (the old spelling re-read).
|
|
887
|
+
|
|
888
|
+
### The engine's capability set names the wake leg — `task.wake` joins `ENGINE_CAPABILITIES`, so a shell gates on a word before it starts a session on its parked notifications (#4106; behavior change: none; narrowing: none; widening: none; hard-breaking: none for a consumer that gates on a word, one for a consumer that pins the whole set (it reads one more member); closed sets: `EngineCapability` +1 member; exports: none; wire: none) (@cli @client-core @server)
|
|
889
|
+
|
|
890
|
+
- **Now.** `ENGINE_CAPABILITIES` (`src/core/engine-capabilities.ts`) carries a fifth word, `task.wake`: the engine reads `TaskSpec.wake` (`TASK_WAKE_SOURCES`, one word, `parked_notifications`), so `runTask({ objective: "", wake: "parked_notifications" })` opens a leg on the notifications parked for the session's occupant and on nothing else, and a wake with nothing handed over ends `config.nothing_to_wake` before any model request. The key itself is 7.67.0's; before this word a shell could tell an engine that reads it only by its version or by trying it (an engine that does not read it answers `config.spec_unknown_key`).
|
|
891
|
+
- **Open to veto — hard-breaking: none; the one break is a pin of the whole set (@cli @client-core @server).** A consumer that gates on a word reads the new one and nothing else changes. A consumer that pins `ENGINE_CAPABILITIES` or `EngineCapability` as a whole (an equality on the array, an exhaustive switch over the type) sees one more member: this repository's own equality pin (`test/backlog1754-held-values.test.ts`) took the new word; `docs/CLOSED-SETS.md` registers it. A server that projects the words it also wires onto its capability face decides on its own whether to project this one.
|
|
892
|
+
- **Rules.** Added: one word, stated in the headed comment where the other four are, in `docs/CLOSED-SETS.md` (the `EngineCapability` row: Five members) and in the integration guide's capability-words paragraph. Retired: none. No table: a word is a statement read by the consumer.
|
|
893
|
+
- **Pins.** `test/backlog4106-task-wake-capability.test.ts` (W1 the word is in the set and the set is frozen — red on the base; W2 a wake with nothing parked ends `config.nothing_to_wake` with zero model requests, W3 a letter in the durable box opens the wake leg and is shown once — each reads the word first, as a shell would, so each is red on the base at that line); `test/backlog1754-held-values.test.ts` V7 (the whole set, five members) and `test/backlog1753-own-credentials-withheld.test.ts` W9 (the registered members equal the array, unchanged).
|
|
894
|
+
|
|
895
|
+
### The additional working directories' instruction files are read under an explicit opt-in — `instructionFamily.additionalDirectories: true` reads each `TaskSpec.additionalDirectories` entry as a further root of the one instruction-file lane (its own level on every load, after the task root's), and the wiring manifest lists what each holds (#3944; follow-up of #3914 / KL-3914-addroots; behavior change: none unless the family opts in — then each additional directory's `files` and `rulesDir` reach the model as root frames of origin `project_instructions`, and every root frame's opening names "the task root or an additional working directory"; narrowing: none; widening: one, below, open to veto @server @cli; hard-breaking: none; closed sets: none; exports: none; wire: `InstructionFamilyDeclaration.additionalDirectories?` (boolean), `WiringManifest.instructions.additionalDirectories?` (`{ directory, selected }[]`), additive) (@server @cli @client-core @test)
|
|
896
|
+
|
|
897
|
+
hard-breaking: none — the opt-in is new and off by default; the root frame's opening sentence is #3914's, unreleased in the same window.
|
|
898
|
+
|
|
899
|
+
- **Before.** #3914 retired the deployment seat that could have read these files (`loadProjectMemory`, and with it #3788's `ctx.additionalRoots`), and the lane read only the task root's tree: an additional working directory's CLAUDE.md reached no reader at all.
|
|
900
|
+
- **Now (the reference's form).** The opt-in rides the family declaration — one declaration, one door (a non-boolean value refuses `config.invalid_argument`, `TaskSpec.instructionFamily.additionalDirectories`), one reader. Omitted or `false` ⇒ none of it is stat'ed or read (the reference's default: `QQ` (cli316.js:270117 c24020) is empty unless the host's asked-once answer or `CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD` opts in). With it, the roots are the task root, then each declared entry as the hands resolved it (a link read at its target and listed under the request's spelling; an entry that does not resolve skipped; the task root itself, or an entry an earlier one already is, read once). A root's own level is owed on every load (law (i)); its frames ride in root order — the task root's level, then each additional root's in declaration order, then the accessed directories below the task root — the reference's order (the additional-directory loop c30240 runs after the working directory's chain).
|
|
901
|
+
- **What an additional root reads.** Its level's `files` and `rulesDir` (the declaration's first family: the reference's loop reads CLAUDE.md, `.claude/CLAUDE.md`, the rules and CLAUDE.local.md and no second name family); no `alternate` (its condition is the task root's); no level below it (an access is an entry below the task root; the reference's nested chain `pet` (cli316.js:272682 c19311) walks only below the working directory); a rule with `paths:` there only for an accessed file below the task root (the reference's loop keeps the rules without `paths:`, c23650 — so for an additional directory outside the task root the two agree); its includes stay inside it. Membership, withdrawal (a delivered member removed or emptied is withdrawn once), the changed-files skip, the compaction restate and the restore exclusion are the root layer's, per root.
|
|
902
|
+
- **Roots that hold each other — one level order.** The levels read in one order: each root's own level in root order, then the levels below the task root; a file two levels name is read under the first of them, the same order `isMember` / the manifest name a member's root by (before, the queue kept whichever item was queued first while membership named the first root — two answers). A declared directory inside the task root is a root (read on the first load, no access needed) and, once accessed, also a level below the task root (that level's `alternate` then applies; a file both name is read once, as the root's). A root inside the task root's rules directory: its CLAUDE.md is the task root's rule (its includes resolve inside the task root). The task root inside a root (or inside its rules directory): the outer root's level is read after the task root's, the task root's files are not its members (the outer root's rules directory claims what lies under it), its rules with `paths:` apply to accesses below the task root (relative to the outer root).
|
|
903
|
+
- **Bounds.** The load's probe bound is `NESTED_MEMORY_PROBES_PER_LOAD` per root read, pooled (a root heavier than its allowance can defer what follows it to the next load; nothing is lost); the frame budget is shared and unchanged. What a load deferred keeps its level's place in the order at the next load (a task-root file the budget left goes ahead of an additional root's).
|
|
904
|
+
- **Also fixed (the same coordinate).** A root that is the filesystem's root (`/`) holds its members by path — membership, the selection, withdrawal and the changed-files skip read them (the one path helper reads no empty root; the task root `/` had the same hole: its accesses and members were unread).
|
|
905
|
+
- **The manifest.** `wiring_manifest.instructions.additionalDirectories` — present iff the family opts in — one entry per additional root read, `{ directory: <the request's spelling>, selected: <that level's members, relative to it> }`; `selected` (the task root's) is unchanged.
|
|
906
|
+
- **Widening (open to veto @server @cli).** A request that opts in puts more repository text in front of the model: each additional directory's CLAUDE.md-level files and rules without `paths:`, under the Read tool's admission (a refusal is told once per session as "an instruction file of an additional working directory").
|
|
907
|
+
- **Text.** The root frame's opening says "an instruction file of the task root or of an additional working directory" (the summarizer's and the classifier's view head: "instruction file of the task root or an additional working directory"); the `TaskSpec.instructionFamily` doc's stale "at the task root itself only the matching rules load" is corrected (since #3914 a root rule without `paths:` loads with the root).
|
|
908
|
+
- **Known limits.** KL-3914-addroots now states the narrow mouth: no level below an additional root, no second family there, its rules with `paths:` only for files below the task root, a directory above the task root read after it, no mid-leg change of the set (the reference re-reads on `/add-dir`, `directory_added` cli316.js:288970 c14439), a read-state key under a spelling of the root that no longer resolves there is no root's member.
|
|
909
|
+
- **Rules.** The root set gains one source (the declared directories, under one opt-in bit on the existing declaration); readers, origins, echo words, notice codes, closed sets: none added. The second family stays the task root's (one predicate, `r === 0`); the include bound and a rule's `paths:` read relative to the root a file is read under (the same rules, one coordinate); the probe bound one allowance per root; one level order read by both the queue and membership (it was two).
|
|
910
|
+
- **Pins.** `test/backlog3944-additional-roots-lane.test.ts` — G1 opt-in omitted / `false`: not stat'ed, not sent, manifest `{ selected }` only · G2 opt-in: the root frame after the task root's, the opening, echoes, the manifest entry under the request's spelling, not sent twice · G3 two roots: membership by root, no level below, a removed name and a removed rule withdrawn per root, the selection per root · G4 order task root → declaration order → accessed levels · G5 (AST) the opt-in is a member of the one family declaration, no second seat, the lane built at one site · G7 a root inside the task root (root first, the accessed level's second family after, nothing twice) · G8 the task root inside a root · G9 a linked root, an unresolvable one, the task root declared again · G10 the first family only outside the task root · G11 includes inside the root · G13 the probe bound one allowance per root (pooled) · G14 the notice's words and the door · G15 an edit sent whole, the changed-files frame does not name it · G16 a deferred file keeps its level's place · G17 a file two levels name read under the first (both containments) · G18 the filesystem root as a root.
|
|
911
|
+
- **handoff:diff / gate:wire (bf642a984..this).** `handoff:diff`: export surface unchanged. `gate:wire`: 0 BREAKING; type keys `InstructionFamilyDeclaration` +`additionalDirectories?`, `WiringManifest` +`instructions.additionalDirectories?` (its entries' `directory` / `selected` required inside the optional member).
|
|
912
|
+
- **Consumers.**
|
|
913
|
+
- @server (S-1989, re-read): #3788's `ctx.additionalRoots` is gone; to read the additional directories' instruction files, set `instructionFamily.additionalDirectories: true` in the family mapping when the deployment's knob says so (the reference's `CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD` or its asked-once answer) — the engine reads them; nothing to load in the server.
|
|
914
|
+
- @cli: map the add-dir opt-in (the env word, the setting, the asked-once answer) to the family's `additionalDirectories` beside `--add-dir` → `TaskSpec.additionalDirectories`; `/memory` reads `instructions.additionalDirectories[]` beside `selected[]`.
|
|
915
|
+
- @client-core: project `instructions.additionalDirectories` if it shows the manifest's instruction files.
|
|
916
|
+
- @test: black-box roots G1 / G2 / G3 / G9 / G15 — off by default; with the opt-in the additional directory's CLAUDE.md on the first request after the task root's, not in the system prompt, listed in the manifest under the declared spelling; removed ⇒ withdrawn once.
|
|
917
|
+
|
|
918
|
+
### The nested instruction lane asks the environment to open a member once and hands the answer to the reader — a first open that was refused or threw is no longer followed by a second open whose answer is read instead (#4209; behavior narrowing, open to veto)
|
|
919
|
+
|
|
920
|
+
- **Rule.** An open's answer is read once, by one reading (`BoundOpening`, `src/tools/fs/bound-open.ts`): `bound` ⇒ the object; `unbindable` (no `openBound`, or `not_supported`) ⇒ by name, judged again; `refused` ⇒ not read. A carrier that holds a file before it judges it hands that answer to the reader (`readText`'s `opened`, which replaces its `bound`; `objectOfOpening` is the tail `readObjectAt` shares) instead of letting the reader open again. The nested instruction lane (`probeFile`) was that carrier, and for every answer that was not held it called `readText` without the object — which asked the environment to open a second time.
|
|
921
|
+
- **Before.** A member whose first open the environment refused (any code but `not_supported`) or threw on was opened a second time by the reader: if that second open held, the member was read from it and delivered; if it refused again the answer was the reader's (a link on the way ⇒ the withheld notice `target_moved`; any other code ⇒ the operator's "the environment could not read it", or nothing when the file was gone); a second throw was the load's catch.
|
|
922
|
+
- **Now.** The environment is asked once. A first open that was refused or threw stands: the member is not read and is told as before — the same sentence, the same channel, per code. Held, and `not_supported` / no `openBound` (by path, KL-3889-d), read as before.
|
|
923
|
+
- **Narrowing (open to veto @server @cli).** Only the shape "first open refused or thrown, second open held" changes its result: the member was read from the second open's object and delivered; it is now not read, and the first answer is told. Every other answer reads byte for byte as before.
|
|
924
|
+
- **Not changed.** The wire; every sentence, notice code and withheld reason; the held read; the by-path read on an environment that cannot hold; the judge; the Read tool's, the attachment's and the compaction restore's own reads.
|
|
925
|
+
- **Rules.** −1: the lane's second, implicit reading of the open (the reader's own re-open) retires; one answer, one reading. Special cases: 0.
|
|
926
|
+
- **Pins.** `test/backlog3889-nested-memory-lane.test.ts` P28c-1 … P28c-5: held · cannot hold · refused × three codes (each told where it was told before) · thrown · a first open refused or thrown while a second would hold ⇒ not read, the environment asked once (the only cell red before the change). Mutations: the lane hands the reader no answer ⇒ P28c-5 red; a refusal read by name in the reader ⇒ P28c-3 and P28c-5 red; a cannot-hold answer refused ⇒ P28c-2 and P28 red; a throw turned into a cannot-hold answer ⇒ P28c-5 (thrown) red.
|
|
927
|
+
- **Consumers.** @server @cli: nothing to do.
|
|
928
|
+
|
|
929
|
+
### A second Read of a file carries its contents on the request, as the first did — only an attachment frame ever stands for an earlier copy; the Read's "unchanged since you last read it" and "already in your context" sentences retire (#4180; ruled C-R1428 and its addendum 1; behavior change: model-visible widening — a repeated Read's request copy is its contents; narrowing: one, the capability word below, open to veto @cli @client-core; widening: one, below, open to veto @server @cli @client-core @test; hard-breaking: none; closed sets: none; exports: none added, none removed at the root (`seededFileUnchangedReminder`, a non-root helper of `src/tools/fs/fs-shared.ts`, is removed: it had no reader left); wire: none) (@server @cli @client-core @test)
|
|
930
|
+
|
|
931
|
+
- **Before.** The request's copy projection rendered a later `Read` result of contents an earlier copy in the same request carried as `[<path>: unchanged since you last read it (lines a-b of n); content omitted to save context]`, and a whole `Read` of a file the leg seeded into its system prompt as `This file is already in your context … Use that content instead of re-reading.` The run's `tool_end` handed out the contents all the while, so the event face and the model face said two different things, and the model that asked for the file again got one line (measured on 7.67.0's tree: a 7 036-byte file read twice — both `tool_end`s 7 240 characters, the request after the second carried 7 240 + 92).
|
|
932
|
+
- **Now.** One judgment in `CopyWalk.decide` (`src/core/request-copy-projection.ts`): only an attachment frame (a `path` attachment, a compaction's restore re-read) stands for an earlier copy. A `Read` result is its contents on every request — a second `Read` of the same window, a `Read` after an `@`-mention of the file, a `Read` of a seeded file — or, as before, the pointer sentence when the request's core cannot hold it (the budget arm). A `Read` still heads a chain, so a later frame of the same contents is the frame's "not repeated" sentence (the attachment half is unchanged). The reference made the same move (cli321.js:296261 "Removed the Read tool's"); a re-read there is only counted (cli321.js:288062 "tengu_file_read_reread").
|
|
933
|
+
- **History.** A transcript an earlier release stored with one of the two sentences in place of a `Read`'s contents is read as before: the sentence is an earlier engine's stub, sent as the pointer sentence, standing for no copy (the next copy of the file in the request carries the contents). The recognizer keeps the two sentences' text and is their only home; no engine writes them.
|
|
934
|
+
- **Who needs it.** The model: it re-reads a file to see it, and now sees it; a host or a person reading the `tool_end` sees what the model saw.
|
|
935
|
+
- **Who is harmed.** A long session that re-reads a large file again and again: each re-read costs its contents where it cost one sentence. Measured (`test/backlog4180-read-dedup-at-request.test.ts` Q1 and the M readings): the same 7 KB file read five times — the request after the fifth carries 5 × 7 240 characters where it carried 7 240 + 4 × 92, so 28 592 characters more (≈ 9 500 tokens at three characters a token). Near the compaction line (Q1: one run reads the same 7 KB file ten times on a 24 000-token window, `reserveTokens` 4 000): on 7.67.0's tree no compaction ran in the eleven requests and the last carried 8 118 characters (one copy, nine sentences); on this tree a compaction ran before the eighth request — the seventh carried 43 482 characters (six copies) — and the session paid one summarization request.
|
|
936
|
+
- **Compensation (measured).** The compaction takes the older `Read` copies out like any tool result: the request after it carried 15 157 characters (two copies in the kept tail and the summary), and the tenth read's request 36 883. The stale-result clear now saves a repeated `Read`'s contents where it saved nothing (`backlog3904-read-state-s1b` G4b: clearing the earlier of two identical `Read`s runs; at #3904 it was judged to save nothing); in Q1's geometry it did not run on either tree (no result cleared).
|
|
937
|
+
- **Narrowing (open to veto @cli @client-core): `user_attachments.read_state`.** The word stays and its promise narrows to the attachment half: a request de-duplicates a `path` attachment's copies and only those; a `Read` result is its contents every time. The word's one statement is its entry in `src/core/engine-capabilities.ts` (`docs/INTEGRATION-CORE.md` §4.6v′, the outcome type and `docs/CLOSED-SETS.md` point there). The statement also says what the attachment half always did and a consumer may now rely on: the earlier copy a "not repeated" sentence stands beside is in that same request — never one a compaction summarized, the stale-result clear emptied, the trim left out or an earlier request carried. **A consumer that read the word as "a repeated `Read` costs no context" reads it again.**
|
|
938
|
+
- **Widening (open to veto @server @cli @client-core @test).** The model reads the contents of every `Read` it makes: a host that sized its own context accounting on the old sentence sees larger requests for sessions that re-read.
|
|
939
|
+
- **Also (the lane's measure).** A folded result (two recorded results of one call, which the tool-turn law sends as one message) is measured as that one message — both segments and the separator — on its first recorded result. It was measured as the two results apart (the separator uncounted) whenever the projection left the fold as written, which is now also the case for two identical `Read` results of one call (`backlog4180` P7).
|
|
940
|
+
- **Not changed.** The attachment frames' three sentences and when each is sent; the pointer sentence and the budget arm; the read state (the edit gate's witness); the stored transcript (every delivery was already stored whole); the `tool_end` of a `Read`; the `attached` outcome word (`src/tools/fs/fs-read.ts` `READ_OUTCOMES`, its sentence now says a `Read` is sent as its contents and a frame may be the sentence).
|
|
941
|
+
- **Rules.** One fewer: a `Read` copy has two renderings (contents, pointer) where it had four (contents, not repeated, seeded, pointer); the "not repeated" and "seeded" arms are the frame's only. Added: none.
|
|
942
|
+
- **Pins.** `test/backlog4180-read-dedup-at-request.test.ts` — R1 a 7 KB file read twice (both copies whole on the request, each equal to its `tool_end`), R2 five reads, R3 a re-read after the first copy was summarized (control), R4 a frame after a `Read` is the "not repeated" sentence (the attachment half stays), R5 a `Read` after an `@`-mention carries the contents, R6 a frame whose earlier frame a compaction summarized carries the contents, R7 the same after a `Read` head, S1 a real file memory backend whose `MEMORY.md` the run seeds (its `Read` carries the contents, a frame after it is the seeded sentence, a second `Read` the contents; changed on disk after the run's system prompt, the frame and the `Read` carry the new contents and the next run's frame is not the seeded sentence), P1–P8 the projection (two `Read`s and a frame; a notebook; a seeded file's `Read` and frame; the budget pointer; a stored stub; a frame after a cleared `Read`; a fold of two identical `Read`s measured whole; a `Read` on a seeded file's chain leaves the seed its head), Q1 the readings near the compaction line. Re-pinned to the new reading (each listed in the car's receipt): `backlog3904-read-state-s1b` (W1, S1, B1 and R1-law and R1-split now read a frame as the later copy, B2, SEED1, R1-spelling, G4, + G4b), `backlog1311a-read-layer` (two cells), `read-seeded-dedup`, `hands-fs` (RB-277 guard), `read-notebook-projection` (B10, B14, B17), `backlog1311a2-attachment-frame` (two cells now read the read gate off an `Edit`), `rb197b-attachment-exclusions-e2e` (the kept-tail read state read off an `Edit`), `backlog894-readstate-after-shrink` (negative control), `design319-reminder-mark` (A2, D3), `backlog1063-reminder-declaration-consistency` (one mint site fewer).
|
|
943
|
+
- **Consumers.** @server: a session's requests grow by a re-read's contents; nothing to change unless it sized on the sentence. @cli / @client-core: the capability word's narrowed promise (above); a card that matched the Read sentence reads the contents. @test: a black-box second `Read` of an unchanged file finds the contents in the next model request, and a second `@`-mention still finds the frame's sentence.
|
|
944
|
+
|
|
945
|
+
### A session whose head cannot be read has its addressed file history deleted by the history store itself, and such a deletion is refused before it deletes anything when a store lacks its leg's `unreadable` verb (#3996 x #4000, met on the stage; behavior change: one optional store member, two root exports, one contract-kit case; narrowing: a host's own history store without the member, and a host coordinator the kit now reads for the missing-verb case (open to veto); widening: none; hard-breaking: none; exports: +`readFileHistoryScope`, +`sessionIdHasScopeTail`, +type `FileHistoryScopeParts`, none removed; wire: none) (@server)
|
|
946
|
+
|
|
947
|
+
- **Before (the stage, once #3996 named history by its address).** A session's history sits at `fileHistoryScopeOf(id, occupant)` = `id~<tail>`, the tail a digest of the occupant token. For a session whose head cannot be read (#4000's operator deletion) the token cannot be read — `headOf` refuses, and the bundled file store leaves the row out of its session list — so step 11 could name no scope: a coordinator written from the table rejected there, after steps 1–10 had run (half-deleted), or left the history on disk (measured on the file family: the run's boundary and tracked set under `id~<tail>` stayed). For a store without a leg's `unreadable` verb the table and the kit said nothing; a coordinator built on the table rejected when it reached that leg, after the earlier legs ran.
|
|
948
|
+
- **Now.** `FileHistoryStore.deleteUnreadableSessionHistory(sessionId)` (optional) erases every scope the store holds whose `readFileHistoryScope(scope).sessionId` is the id — the bare id and each occupant's address; never an id the session id is a prefix of (`s1x`, `s1-x`) or one carrying `~` of its own (`s1~a`) — and answers the total `deleteBySession` would; a repeat answers 0. The in-memory and file stores carry it; the file store reads each scope back from its tracked set (the directory name is a one-way hash), and a directory it cannot read back — a tracked set that names no scope or another directory's, one missing beside a boundary, a directory it cannot stat — fails the call before anything is erased (a directory with only an attempt's staging is no scope); a repeat that finds no scope runs the blob pass again, as `deleteBySession`'s own repeat does. `RETENTION_CASCADE` step 11 names it as its `unreadable` verb; step 10, for such a session, carries each fork that records a source occupant from the address its own row records (every scope under the id is about to go).
|
|
949
|
+
- **One rule for the three `unreadable` verbs.** A deletion of a session whose head cannot be read first checks that every store whose leg names an `unreadable` verb carries it (`SessionStore.deleteUnreadable`, `CheckpointStore.deleteUnreadableSessionCheckpoints`, `FileHistoryStore.deleteUnreadableSessionHistory`) and is refused before its first destructive call, naming the store and the verb, when one does not: nothing deleted, nothing half-deleted. Written once, on the table's `unreadable` mechanism (`src/core/retention.ts`).
|
|
950
|
+
- **The grammar and its inverse.** `scope = sessionId | sessionId "~" tail`, `tail` exactly 22 characters of `[A-Za-z0-9_-]` ending the scope; a session id may carry `~`, `%` and `_`; read from the end — the scopes the engine mints under the address epoch are unambiguous (an id that ends like an address has no bare scope, #3996's refusal); the one ambiguous shape is a pre-epoch bare scope of such an id, which reads as an address of the shorter id — a known limit, KL-3996-a (measured, `test/backlog3992-file-history-occupant-address.test.ts` G4: the shorter id's unreadable deletion erases it; a host censuses its ids with `sessionIdHasScopeTail` and re-ids them). `readFileHistoryScope(scope)` (package root) answers `{ sessionId, tail? }`, or `undefined` for a string no session's scope is; `sessionIdHasScopeTail(id)` (package root) is the one predicate behind #3996's refusal and this reading. Written on `fileHistoryScopeOf` and in INTEGRATION-CORE (the file-history address section).
|
|
951
|
+
- **Kit.** `sessionDeletionCascadeContract`: the recording session list leaves a session whose head cannot be read out, as the bundled file store does; one case added — such a session over stores of which one lacks its leg's `unreadable` verb (each in turn) is refused before any destructive call, and a coordinator that offers the operator's deletion names the store and the verb. `fileHistoryStoreContract`: one case for the new member when a store carries it (the prefix ids above keep their history).
|
|
952
|
+
- **Narrowing (open to veto) @server.** (1) A host's own `FileHistoryStore` that does not carry the member: the deletion of a session whose head cannot be read, through a coordinator that follows the table, is refused up front (before: the history legs ran by the bare id and the deletion completed; since #3996 the bare id no longer holds an addressed session's history, so that completion left it on disk). To implement it: erase every scope whose `readFileHistoryScope(scope)?.sessionId` equals the id (or file that `sessionId` in a column and select by it) — not `LIKE 'id~%'`: the tail carries `_` and `-` and an id may carry `%` and `_`. (2) A host coordinator run through the kit meets the new case: one that reaches a missing verb mid-cascade fails it.
|
|
953
|
+
- **Rules.** Readings of "a session whose head cannot be read, over a store without its leg's `unreadable` verb" 0 (each coordinator's own mid-cascade rejection) → 1 (refused before the first destructive call); readings of "which session does this history scope belong to" 0 → 1 (`readFileHistoryScope`, the inverse written beside the one mint; no store matches a prefix). Added: one optional store member, one pure function, one kit case. Retired: the half-deleted shape for that session.
|
|
954
|
+
- **Consumers.** @server: the member on a host's own history store (above), and any history delete keyed by session reading scopes back with `readFileHistoryScope` instead of the bare id; its stock of caller-chosen session ids from releases up to 7.125.x is where KL-3996-a's shape can stand — census it with `sessionIdHasScopeTail` before an unreadable-head deletion runs over it. @cli @client-core: nothing.
|
|
955
|
+
- **Pins.** `test/backlog4000-cascade-missing-stores.test.ts` C2 (real file store, header dropped: no file names the session after the cascade but what a leg keeps) · C2b ×3 (each store without its leg's verb: refused naming store and verb, nothing on disk moves) · C2c (the real store's list and the kit's recording list both leave the unreadable session out) · C1 / C6 (the readable census and the kit, history by address) · `test/backlog3005-session-deletion-cascade.test.ts` P1 / P1b / P4 and P3 `verbs-unchecked` (case `unreadableVerbs`) · `test/backlog4018-roster-background-rows.test.ts` K1 · `test/backlog3992-file-history-occupant-address.test.ts` G1–G3 (mint then read back is the identity for ids with `~` `%` `_`; `s1` / `s1x` / `s1~a`; strings no scope is) and G4 (MEASURED, not a law: the pre-epoch shape above, on both stores) · `fileHistoryStoreContract` "deleteUnreadableSessionHistory" (in-memory and file) · `test/file-history-file-store.test.ts` U1 (tracked set missing beside a boundary: refused, nothing erased) · U2 (staging only: not a scope) · U3 (a failed blob pass, then repeats: loud while it fails, then 0).
|
|
956
|
+
|
|
957
|
+
### A notice seat that holds no function no longer fails the run of a deployment that declares no instruction family (#4179 met the malformed-seat contract on the stage; behavior change: a fix — the run completes; narrowing: none; widening: none; hard-breaking: none; exports: none; wire: none)
|
|
958
|
+
|
|
959
|
+
- **Before (the stage).** `RunnerDeps.onNotice` holding `null` or a string — the malformed seat `deliverEngineNotice` falls back to `console.warn` for — failed every run at prepare once `instructions.family_undeclared` was due: its once-per-receiver record keyed a `WeakMap` on the seat, which refuses a non-object key.
|
|
960
|
+
- **Now.** A seat that is not a function is no receiver, as the delivery throat reads it: the record's unit is the process (the same record as no seat wired), and the line reaches the console once, like every notice of a malformed seat. `noticeOncePerSink`, `src/core/runner/announce-once-ledger.ts`.
|
|
961
|
+
- **Pins.** `test/onnotice-structured-channel.test.ts` station ① / ② malformed seat (both completed again) · `test/backlog4179-instruction-family-undeclared.test.ts` U9 (null and a string, two Runners, three runs: all completed, one console line).
|
|
962
|
+
|
|
963
|
+
### A tool card too large for its `tool_end` leaves its skeleton — `tool_end.structuredSkeleton`: the card's `type` plus the members its kind declares fixed-length; the four file cards carry `contentSha256`, the engine's own hash of the text it wrote (#4200; cli EXTNOTE166 ⑨ / L-3007; behavior change: one optional frame member, one optional member on the four file cards; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed (`createStructuredProjector` returns what it returned); wire: `tool_end.structuredSkeleton?`, `structured.contentSha256` on `create` / `update` / `edit` / `notebook-edit`) (@cli @client-core @server @test)
|
|
964
|
+
|
|
965
|
+
- **⚠ `contentSha256` is NOT a digest of the bytes on disk (@cli @client-core @server).** It is the sha-256 of the file's text as the engine reads it back after the write — decoded, a leading BOM dropped, CRLF as LF. On a file with a BOM or CRLF endings it differs from `sha256sum` of the file; compare it with the sha-256 of the file's text read back the same way, never with the raw bytes, or every such file reads as changed after the agent wrote it.
|
|
966
|
+
- **Rule.** A card the frame cannot carry rides as its skeleton: `type` plus the members its kind declares fixed-length, copied as the tool wrote them (a dotted name reaches into a nested card). The members are declared once per kind in `CARD_SKELETONS` beside `ENGINE_CARDS` (`src/core/tool-catalog-entries.ts`). The table covers every engine card: a card with no row, or a row naming no card, stops the catalog from loading. A caller-declared card (`renderHints.resultCards`) has no row, so its skeleton is `{ type }`. The bound (`cardOnFrame`, `src/core/tool-card-bound.ts`) knows no kind. A skeleton the frame cannot carry either rides as nothing; it is never cut, because a shortened path names another file.
|
|
967
|
+
- **Before.** A card over 65 536 JSON characters was left off `tool_end.structured` whole, and the frame said nothing else. A host could not tell "this tool mints no card" from "its card was too large". For a file write it could not learn which file was written or what went into it. A shell kept its own fallback reading of the call's input for that case (cli L-3007).
|
|
968
|
+
- **Now.** `structured` is unchanged, and so are the three written sentences about it ("omitted whole, never half"). An over-bound card's skeleton rides the new member `tool_end.structuredSkeleton`. The two members are never present together, so which one is present says whether the card is whole; there is no marker member. The `create`, `update`, `edit` and `notebook-edit` cards carry `contentSha256` whether they ride whole or as a skeleton. It is the hex sha-256 of the file's **text as the engine reads it back after the write** (decoded, a leading BOM dropped, CRLF as LF). The same value is written to the read state and to the card. It equals the sha-256 of the bytes on disk for a UTF-8 file with LF endings and no BOM, and differs for a file with a BOM or CRLF endings.
|
|
969
|
+
- **Not changed.** The bound and its measure. The plan-exit tool's refusal before it asks (#4120) and the plan pause's reading of `structured`. Every member of a card that fits, except the one added `contentSha256` on the four file cards. `createStructuredProjector`'s return value.
|
|
970
|
+
- **Rules.** Retired: each consumer's private reading of an absent card. One rule declared per kind replaces it, and the bound gains no kind-specific branch. Added: the skeleton rule and its one roster. The file cards' digest is the read state's own value, not a second hash.
|
|
971
|
+
- **Pins.** `test/backlog4200-card-skeleton.test.ts`, three groups:
|
|
972
|
+
- Real Runner, production defaults: Write create / update, Edit single / batch and NotebookEdit each go over the bound. The skeleton's key set is exact and its `contentSha256` equals the disk sha of the LF file. An in-bound control rides whole with no skeleton.
|
|
973
|
+
- Projection: one over-bound card per engine kind gives exactly its declared skeleton, and the published projector returns nothing for it. A caller card gives `{ type }`. A skeleton that does not fit rides as nothing, never cut. The roster equals the card set.
|
|
974
|
+
- `contentSha256` equals the read-state hash for LF, CRLF and BOM files written by Write and Edit, and for NotebookEdit. Whether it equals the disk-bytes sha is recorded per file.
|
|
975
|
+
- **Consumers.**
|
|
976
|
+
- @cli: L-3007 can read `structuredSkeleton.filePath` / `contentSha256` instead of the call's input. Compare `contentSha256` with the sha-256 of the disk text read back the way the engine's Read reads it, not with the bytes.
|
|
977
|
+
- @client-core: `executedFileCardOf` can take a whole card or a skeleton by the same members.
|
|
978
|
+
- @server: passes `tool_end` through; the new key is `structuredSkeleton`, to be redacted like `structured`.
|
|
979
|
+
- **Wire.** Additive: `tool_end.structuredSkeleton?: { type: string } & Record<string, unknown>`, and the card member `contentSha256` (string) on the four file cards.
|
|
980
|
+
|
|
981
|
+
### A person's rule store that cannot be read says which partition, why, and what to repair — in the run's notice (told once per session per fault set), in the question the call is put to a person with (so in what a headless model reads), and on the card's "why it cannot be cleared" half (#2355; behavior change: one notice sentence, one question sentence and the card's shadow half reworded, the notice told once per session per fault set instead of once per run; narrowing: none; widening: the notice is told fewer times, below, open to veto @server @cli @client-core @test; hard-breaking: none; closed sets: one added, `RuleStoreFaultReason` (8 words); exports: 4 added (`RULE_STORE_FAULT_REASONS`, `RuleStoreFault`, `RuleStoreFaultReason`, `RuleStoreSource`), none removed; wire: additive, below) (@server @cli @client-core @test)
|
|
982
|
+
|
|
983
|
+
- **Before.** A store that could not be read (a directory with no permissions, one writable by others or another account's, a marker file nobody may read, a damaged bucket, a bucket a newer engine wrote) was told as `your persisted permission rules could not be read … none of their allow rows … applies …` with `detail: { at, sessionId, runId }`: no partition (unless the view happened to carry one), no cause, no repair; the call's question read `the person's standing permission rules could not be read — whether one of them forbids this call is unknown, so it needs a real approval` (what a headless model reads as the result), and the card's allow-shadow half read one three-way sentence (`this deployment's organization policy, a partition of the person's standing rules, or this call against a deny or ask rule`) whichever it was. The cause existed — the file store's refusals carry `code` and `errno`, and `unsafeDirFault` knew "owned by another user" from "writable by group or others" — but `effective()` flattened every refusal to text and the readers dropped it. The notice was minted once per RUN, so a session's every message (every run) said the same long paragraph again. Measured on the base (real Runner, real file store, one run with two calls, three shapes × `default` / `bypassPermissions`): the notice is minted once per run in 6 of 6 cells; the trace frame `permission.rule_store_unreadable` and the store's own `onError` disclosure line (it carries the path) are written once per READ, five times in that run.
|
|
984
|
+
- **Now.** One reader turns a refusal's kind into a closed word: `ruleStoreFaultReasonOf(code, errno, unsafe)` — `not_readable` (`bucket_unreadable` with `EACCES` / `EPERM`), `read_failed` (any other failed read), `unsafe_permissions` (`bucket_damaged`, directory writable by group or others), `owner_mismatch` (`bucket_damaged`, directory another account's), `damaged` (every other `bucket_damaged` — and a row any backend served whose behavior word is outside deny / ask / allow: its words, `the stored rules, or what holds them, are damaged or not in the form this engine expects` / `Restore a good copy of them, or clear them (clearing them gives up every saved rule).`, name no file or folder, so they are true of a host's own store too and of the file store's other `damaged` refusals (a link or a non-directory in the store directory's place, an adoption marker that is no marker); the four file-form words — `not_readable`, `read_failed`, `unsafe_permissions`, `owner_mismatch` — are the file store's, and a store that is no file states only the others), `later_grammar`, `no_answer` (the read did not answer in time) and `unclassified` (a store that states none of these). The refinement `unsafe: "permissions" | "owner"` is a new own member of the file store's refusal beside `code` and `errno`, on the two unsafe-directory arms only; the adoption marker's refusal (a plain `Error` with no code until now — the first thing a mode-000 directory fails on) carries the bucket's codes. The view says it: `EffectivePermissionRules.unreadableFaults` (one `{ sources, reason }` per failed partition — the durable partition covers `user` and `project` — and `unreadableSources` is its flat list, one mint) and `PersonRowsRead.unreadableFaults`; a read that failed whole (the store door threw, or it did not answer) is one fault over every partition it took out. **The notice** `permission.rule_store_unavailable` carries `detail: { at, sessionId, runId, faults: [{ sources, reason }] }` and says `Your saved permission rules (user and project) could not be read when this run started: <clause>. ` followed by the consequence paragraph it always carried (none of their allow rows applies, the deny and ask rows most recently read still do and a deny row you removed keeps refusing until a read succeeds; every call is asked about rather than allowed, whatever its tool), then `<the repair> Every later check reads them again.` — one clause and one repair per word (e.g. `the folder that holds them belongs to another account` / `Make the account that runs this agent the owner of that folder again, or restore a copy of it that account owns.`); no path and no permission mode (the class is one no mode releases). A way out that says "that folder" ends by saying where it is told — `The operator's error log says which folder.` (`not_readable`, `unsafe_permissions`, `owner_mismatch`); `read_failed`, `damaged` and `unclassified` end `The operator's error log has the details.`; each pointer is said once however many faults share it, and `read_failed`'s way out no longer ends "then carry on" (that is the model's line, in the question). The operator's log the pointers name is the trace frame `permission.rule_store_unreadable` (it carries the store's refusal text) and the file store's own `onError` (it carries where): a deployment that wires neither has no log to point at. **The question** (store form) reads `your saved permission rules (user and project) could not be read (<clause>), so whether one of them forbids this call is unknown and it needs a person's approval — if nobody can be asked, tell the person what is wrong with the rules store and carry on with the parts of the task that do not need this call`, from the cause the lane's store-form answer now carries (`PersistedRuleUnreadable.faults`, read through a closed-set screen; a lane that states none keeps the generic clause). **The card's shadow half** for class `rule_unreadable` says the one form the ask's origin names — `this deployment's organization policy cannot currently be read …` / `the person's standing permission rules cannot currently be read …` / `this call cannot be read against the person's standing deny and ask rules …` — and keeps the three-way `SHADOW_REASON.rule_unreadable` row for an ask that does not say which. **Told once per session per fault set:** through the session's announced record (`oncePerSession`, the prepare-time seat every once-per-session notice uses; digest = `announceDigest(code, <the fault set, sorted>)`), so a session's later runs over the same unreadable store say nothing, a different cause or partition is told again, and a run on a session with no record to read says it once per run. Every approval card keeps its own cause each time.
|
|
985
|
+
- **Widening (open to veto @server @cli @client-core @test).** The notice is told fewer times: a session's second and later runs over the same fault set (same partitions, same reason) no longer repeat it, even after a read that succeeded in between (told once per session per fault set — a delegated agent's child runs on a session of its own, so each child tells it once itself), and a fault first met at a call (the prepare read was fine) is not written into the record that outlives the run, so it may be told once more by the next run. A consumer that counted one notice per run for this code reads one per session per fault set; the question on every card and the trace line on every read are unchanged.
|
|
986
|
+
- **Not changed.** The direction: a rule store nobody can read is still a real-approval ask of class `rule_unreadable` in every mode (`bypassPermissions` keeps it, a headless run refuses it); which rows apply while it cannot be read; the call form's question (`this call could not be read against the person's standing deny/ask rules (<reason>)`); the org-unavailable question; the store's refusal sentences and their no-location law (the store's disclosure line still carries where); the trace frame and the `onError` line once per read (a shell that routes either to its stderr prints one line per read — that is the shell's, not the notice's).
|
|
987
|
+
- **Rules.** One reader (`ruleStoreFaultReasonOf`) and one table of clauses and repairs where the engine had prose in four places; the cause is read from `code` / `errno` / `unsafe` and never from a sentence. Retired: KL-2229-c (the cause not on the card or the notice); narrowed: KL-1702-a and KL-1653-h (each says what remains); added: KL-2355-a (a read that failed whole — the store's door threw — names every partition it took out, a session partition included, beside the durable store's cause; reachable only for a host that wires a session rule store).
|
|
988
|
+
- **Wire.** Additive: `permission.rule_store_unavailable.detail.faults`; `EffectivePermissionRules.unreadableFaults?` and `PersonRowsRead.unreadableFaults?`; `PersistedRuleUnreadable.faults?`; four root exports (above); `SHADOW_REASON` is unchanged (the per-form sentences are minted beside it); `allowRuleShadowSentence` / `cardDisclosureOf` take two optional members (`origin`, `ruleUnreadable`); the file store's refusal error gains the optional own member `unsafe` and the adoption marker's refusal gains `code` / `errno`.
|
|
989
|
+
- **Pins.** `test/backlog2355-rule-store-unreadable-sentence.test.ts` — ① seven shapes (directory with no permissions, writable by others, another account's, unreadable marker, damaged bucket, bucket of a newer engine, a store that throws) × `default` / `bypassPermissions`, one run with two calls: one notice, `detail.faults`, sentence, the refused calls' text, no path anywhere; the direction is unchanged · ② the card's shadow half per form · ③ the one reader's table and the closed-set screens · ④ the file store's refusals · ⑤ a read that does not answer · ⑥ one session, many runs · ⑦ the words are true of the store that stamped them (a host's own store serving a row it cannot read is not told as a damaged file; only the four file-form reasons name a folder or a file; a located way out ends with where it is told, once).
|
|
990
|
+
- **Consumers.** @cli @client-core: the notice's `detail.faults[].reason` is a closed word to render (`RULE_STORE_FAULT_REASONS`); a shell that wants to show where the rules live knows its own rules directory (no path rides the notice). @server: the same words are on the view for an operator surface; a deployment that counted the notice per run counts it per session per fault set. @test: a black-box read of a run over each unreadable shape finds the partition, the cause and the repair in the notice, and the cause in the headless refusal's text.
|
|
991
|
+
|
|
992
|
+
### A read-only command wrapped in a plain `timeout DURATION` no longer asks once more than the bare command — the read-only reader peels the wrapper the way the reference does (#4157; behavior change: `timeout 30 rg foo .` and its read-only kin run without a question under `default`; narrowing: none; widening: one, below, open to veto @server @cli; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none) (@server @cli @test)
|
|
993
|
+
|
|
994
|
+
- **Before.** Under `default` the bare `rg foo .` ran without a question while `timeout 30 rg foo .` asked once (measured on the real Runner with an approver that refuses: 0 asks against 1, the same for `timeout 30 timeout 5 cat a.txt`). The reader that proves a command read-only (`readOnlyShellVerdict`, `src/core/read-only-shell.ts`) took off a `command` / `builtin` prefix and stopped there, so a `timeout` stayed the program and was in no table.
|
|
995
|
+
- **Now.** After the `command` / `builtin` prefix the reader also takes off a plain `timeout DURATION` — the duration a number with an optional `s` / `m` / `h` / `d` — as often as it stacks, and reads the words that remain as the command, with every check the bare command meets (its table row and flags, its redirections, a glob, the face's `readers`, the cd-beside-git line objection). It is one peel in the reader's own prefix walk; no second wrapper table. A `timeout` that carries an option of its own (`-s KILL`, `-k 2`, `--preserve-status`), a wrapped shell builtin (`[[`, `alias`, `cd`, `help`, `history`, `type`), a missing or malformed duration, and nothing wrapped are not peeled and keep their question — the reference does not read them either.
|
|
996
|
+
- **Widening (open to veto @server @cli).** A read-only command under `timeout DURATION` is now allowed without a question wherever the bare command is: the tool gate's read-only arm, a compound's segment that needs no rule, the read-only shell face and the transcript-directory gate (all ask the one reader). A command that writes, runs another program or carries a redirection that writes asks exactly as it did; the wrapper opens no door the bare command is shut to.
|
|
997
|
+
- **Not changed.** Every table row and flag set, the `command` / `builtin` peel, the question for a `timeout` with options, `noglob` (a program here, as before), what the classify face and the poll-loop face read.
|
|
998
|
+
- **Rules.** Added: none (one more prefix word in the reader's existing prefix walk; the helper that took off `command` / `builtin` is renamed `peelWrapperPrefix` since it now takes off two kinds). Retired: none.
|
|
999
|
+
- **Consumers.** @server @cli: a card or transcript that showed a question for a `timeout`-wrapped read now shows none. @test: the reference reads the same (RT-12, real 2.1.321).
|
|
1000
|
+
- **Pins.** `test/backlog4157-timeout-wrapped-readonly.test.ts` — the bare controls, the wrapped read forms (durations, stacked, after `command`, in a pipe), the non-read-only controls (`rm`, `touch`, `sh -c`, `git commit`), the wrapper opening no door (`rg --pre`, `find -delete`, `sort -o`, a writing redirect), the option forms and wrapped builtins not peeled, the malformed wrappers, the verdict on its own, the face's `readers`, the cd-beside-git objection, and segment-by-segment compounds.
|
|
1001
|
+
|
|
1002
|
+
### The longest principal name the engine keeps is a root export — `PRINCIPAL_NAME_MAX_CHARS` (256 UTF-16 units) stands beside `principalNameOf`, so a host that checks a name at its own door reads the number from the package (#4226; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: `PRINCIPAL_NAME_MAX_CHARS` +1 root constant; wire: none) (@server)
|
|
1003
|
+
|
|
1004
|
+
- **Now.** `PRINCIPAL_NAME_MAX_CHARS` is the constant of `src/core/principal-name.ts` (until now an export of that module only, not of the root, spelled `NAME_MAX_CHARS` there and renamed so the name on the root says whose bound it is — a deep import of `dist/core/principal-name.js` that read `NAME_MAX_CHARS` reads `PRINCIPAL_NAME_MAX_CHARS`), the one bound `principalNameOf` judges by, the task door's refusal sentence names ("a string longer than 256 characters") and the mailbox's name spelling shares; the root now exports it. No value moves and no second constant is made. @server: a pre-check of a principal name can read it instead of a copied 256. Pin: `test/backlog4226-principal-name-bound-root-export.test.ts` (the exported value is the engine's constant; a principal of exactly that many units runs, one more is refused at the task door by the sentence that names it).
|
|
1005
|
+
|
|
1006
|
+
### The model fallback gate's closed words are readable at runtime from the package root — `DEGRADE_REASONS` (the words the policy door's membership test for `RunnerDeps.modelFallback.on` is built from, through `isDegradeReason`), `MODEL_FALLBACK_GATE_WORDS` (the list the door's refusal sentences name) and `DEFAULT_FALLBACK_GATE` (the gate when none is configured) are the frozen arrays the engine itself holds (#4214; behavior change: none; narrowing: none; widening: none; hard-breaking: none; closed sets: no word added or removed (`DegradeReason` is registered in `docs/CLOSED-SETS.md`); exports: `DEGRADE_REASONS`, `MODEL_FALLBACK_GATE_WORDS`, `DEFAULT_FALLBACK_GATE` +3 root constants; wire: none) (@server @cli)
|
|
1007
|
+
|
|
1008
|
+
- **Before.** The root carried only the types `DegradeReason` and `ModelFallbackPolicy`. The words themselves (the reason words, the seven a gate may hold — every reason but `budget` — and the four-word default gate) lived inside `src/core/model-fallbacks.ts`, so a deployment's documentation, or a shell completing and validating `MODEL_FALLBACK_ON`, could only copy a list and keep it in step by hand.
|
|
1009
|
+
- **Now.** Three runtime constants on the root, each the array the engine itself holds — not a copy: `DEGRADE_REASONS` (the eight reason words `TaskResult.degraded.reason` can carry; the door judges a word of `on` by `isDegradeReason`, a set built from this array, and refuses `budget` by name), `MODEL_FALLBACK_GATE_WORDS` (the reason words without `budget`, in the same order: what `modelFallback.on` may hold; the door's refusal sentences list it, its membership test does not read it), `DEFAULT_FALLBACK_GATE` (`rate_limit`, `breaker_open`, `server_error`, `last_resort`: the gate when none is configured). All three are frozen. The two refusal sentences that list the words already built the list from the one array; they now read it under the exported name (`MODEL_FALLBACK_GATE_WORDS`, formerly the module-internal `GATE_WORDS`; `DEGRADE_REASONS` and `DEFAULT_FALLBACK_GATE` were already exports of the module, not of the root) and are word for word what they were.
|
|
1010
|
+
- **Not changed.** Every word, the default gate's content and order, what the policy door accepts and refuses, and the sentences it says.
|
|
1011
|
+
- **Rules.** Added: none (three existing arrays on the root — two under the names they already had, one renamed from `GATE_WORDS` — and that one frozen where it was not before). Retired: none. `DegradeReason` is registered in the closed-set registry.
|
|
1012
|
+
- **Consumers.** @server: the fallback section of the documentation and the config manual can read these symbols instead of a copied list. @cli: completion and validation of `MODEL_FALLBACK_ON` can read `MODEL_FALLBACK_GATE_WORDS`.
|
|
1013
|
+
- **Pins.** `test/backlog4214-fallback-gate-root-exports.test.ts` (G1 each root name is the object the door reads, G2 frozen, G3 the words and their order, G4 the door judges against them, G5 the refusal sentences list exactly the exported words and read word for word as before).
|
|
1014
|
+
|
|
1015
|
+
### An approval card says first, in plain words and in the run's language, why you are asked and whether saving a rule would stop it being asked again — one new optional member, `headline`, on the live card, the parked row and the inbox listing; six engine sentences rewritten for the person who reads them (#4130)
|
|
1016
|
+
|
|
1017
|
+
- **Rule.** A card's first line is a rendering of its class word and nothing else: one catalog row per `ASK_CLASSES` word says why the reader is asked (26 rows; `content_question` — the agent's own question — has nothing to say beyond its card's header and carries none: no row only repeats the card), and the class's own `ruleMayClear` column of `MODE_ARBITRATION` picks one of two remedy rows ("Saving an allow rule for it would stop it being asked again." / "Saving an allow rule would not stop it being asked again."). The class word's row is the folded row of the ask's whole class set, because the only classes a rule may clear (`unstated`, `ordinary`) are the last two in class order — a structural pin holds that, so a third such class or a reorder turns it red. One composer (`src/core/ask-headline.ts`), two mint points: the resolver's chokepoint (`resolveAskFor`, which every request station passes — a caller's own `headline` is replaced) and the park's carry (`askCarryRowMembers`); the listing echoes the row's value through the screen `cardMessage` takes. The line is minted in the run's `promptLocale` (canonical and zh-Hans rows in the catalog); a host that calls `resolveAsk` itself gets the canonical line.
|
|
1018
|
+
- **Before.** The card's first glance was the engine's detailed account (`message`) plus words each surface composed from typed members — internal words on the first line (test [13890] #1: `egress: not annotated`, `is not in the read-only allowlist`, `the coarse shell check rates this command above the tier it lets through`).
|
|
1019
|
+
- **Now.** `AskRequest.headline?: string`, `PendingAction.tool_approval.headline?: string`, `CheckpointSummary.headline?: string` — same name, same position, and on each face the rendering of the class word that face carries (the listing echoes the row). A card re-sent from a parked row carries the line minted at the park, in that parking run's `promptLocale`: a resume does not re-mint it. Where two faces of one question carry different class words, their lines differ with them: an inherited ask its ancestor's approver answered `unavailable` is classed under the station's facts on the live card and as `ancestor` on the row it parks (pre-existing; the row's word is what a parked decision is judged by). **For people to read, never a branch input**: branch on `askClass` / `classes` / `origin` / `mandate` / `ruleOffersAbsence` as before. A host that has its own sentence for a class word decides which of the two its card shows first; the engine does not promise the two say the same thing. **Absent** ⇔ the ask carries no class word (a request a host built, the shared-artifact read confirmation, a row an earlier engine parked, an older engine) or its class word is `content_question`: the engine gave no headline — show `message`; absence is not a claim that there is no reason. A listing value that is not a non-empty string reads as absent. This section does not change `message` (the detailed layer); the #4196 group, below, rewrites some of its sentences (a rule's ask reads `the permission rule "…" requires approval before a call like this runs`, where 7.67.1 read `flagged by permission rule "…"`).
|
|
1020
|
+
- **Not changed.** `message`, `cardMessage` and `askReasons` as this section found them (their words move in the #4196 group, below), every class / origin / mandate word, every decision; no gate, CAS, settlement or resume path reads the new member (row requirement `SHOWN`).
|
|
1021
|
+
- **Hard-breaking.** None. **Narrowing / widening.** None. Old readers: zero change (an optional member they do not read).
|
|
1022
|
+
- **Rules.** Readings of "why am I asked, in plain words" across the stack: one engine table (26 + 2 + joiner rows, fenced to the class set with the one named no-headline class) where each surface composed its own line; a surface that renders `headline` and keeps the words for branching needs no catalog of its own for this line.
|
|
1023
|
+
- **Closed sets.** None (the class set is read, not extended). Catalog: `ENGINE_MESSAGES` + 29 (`core/engine.ask-headline.<class>` × 26, `core/engine.ask-headline-remedy.{rule-clears,rule-cannot}`, `core/engine.ask-headline-join`), each `securityBearing` (its zh-Hans rendering waits for the native-reader review record).
|
|
1024
|
+
- **Wire.** Additive: `AskRequest.headline`, `PendingAction.tool_approval.headline`, `CheckpointSummary.headline`. Consumers @server (relay on the live `tool_approval` card, the `approval_request` card and the ops-queue row; the row store round-trips one more optional member) @client-core (CC-1517: the three faces read it; the per-origin sentence family can render it instead) @cli (render it as the card's first line; `message` behind ctrl+e) @web-client.
|
|
1025
|
+
- **The 26 rows** (canonical / zh-Hans; remedy = the class's `ruleMayClear`: ✓ "Saving an allow rule for it would stop it being asked again." / 为它保存一条允许规则,以后就不会再问。 · ✗ "Saving an allow rule would not stop it being asked again." / 保存允许规则也不能让它以后不再问。; `content_question` carries no headline):
|
|
1026
|
+
|
|
1027
|
+
| class | remedy | canonical | zh-Hans |
|
|
1028
|
+
|---|---|---|---|
|
|
1029
|
+
| `hook` | ✗ | A check this deployment runs before tools wants you to decide this step. | 这个部署在工具运行前设的检查要你来决定这一步。 |
|
|
1030
|
+
| `ancestor` | ✗ | The agent that started this one needs your approval for this step. | 启动这个智能体的上一级智能体需要你批准这一步。 |
|
|
1031
|
+
| `dangerous_delete` | ✗ | This deletes a top-level, home or workspace folder, or a folder that holds one. | 这会删除顶层目录、主目录或工作区目录,或包含它们的目录。 |
|
|
1032
|
+
| `unverifiable_delete` | ✗ | This deletes files, and the command does not say which ones. | 这会删除文件,但命令里没有写明删哪些。 |
|
|
1033
|
+
| `unverifiable_shell_text` | ✗ | The engine cannot read exactly what this command will run, so it cannot be checked first. | 引擎读不出这条命令实际会运行什么,无法事先检查。 |
|
|
1034
|
+
| `transcript_write` | ✗ | This changes the record of this session's own conversation. | 这会改动本会话自己的对话记录。 |
|
|
1035
|
+
| `org_rule` | ✗ | Your organization's policy asks for your approval here. | 你所在组织的策略要求这里由你批准。 |
|
|
1036
|
+
| `rule_unreadable` | ✗ | A permission rule that may cover this could not be read, so you are asked instead. | 一条可能适用于这一步的权限规则读不出来,所以改为问你。 |
|
|
1037
|
+
| `ask_rule` | ✗ | A rule in the settings asks before steps like this. | 设置里有一条规则要求这类步骤先问。 |
|
|
1038
|
+
| `denial_limit_fallback` | ✗ | Automatic approval has turned down as many steps as this session allows, so this one comes to you. | 自动审批拒绝的步骤已达本会话允许的上限,所以这一步交给你。 |
|
|
1039
|
+
| `classifier_no_ruling` | ✗ | Automatic approval could not decide on this step, so it comes to you. | 自动审批没能对这一步作出判断,所以交给你。 |
|
|
1040
|
+
| `requires_user_interaction` | ✗ | This tool says you have to take part in this step. | 这个工具声明这一步必须由你参与。 |
|
|
1041
|
+
| `cross_boundary` | ✗ | This session may run on a device outside the deployment's trusted network, and this step may cross between the two. | 本会话可能运行在部署信任网络之外的设备上,而这一步可能要跨越两者之间。 |
|
|
1042
|
+
| `operator_forced` | ✗ | This deployment marks steps like this as needing your approval. | 这个部署把这类步骤标为需要你批准。 |
|
|
1043
|
+
| `shell_floor` | ✗ | This deployment asks you before every command. | 这个部署要求每条命令都先问你。 |
|
|
1044
|
+
| `target_unresolved` | ✗ | Where this writes cannot be worked out before it runs. | 这一步会写到哪里,运行前无法确定。 |
|
|
1045
|
+
| `probe` | ✗ | This may reach outside this session's folders or be hard to undo, and a check could not rule that out. | 这一步可能超出本会话的目录或难以撤销,检查无法排除这一点。 |
|
|
1046
|
+
| `write_protection_on_disk` | ✗ | This writes to a protected file through a link or another spelling of its path. | 这会经由链接或路径的另一种写法写入受保护的文件。 |
|
|
1047
|
+
| `write_protection` | ✗ | This writes to a protected file, such as the agent's own settings. | 这会写入受保护的文件,比如智能体自己的设置。 |
|
|
1048
|
+
| `tool_marks_mcp` | ✗ | The tool's server marks this as sending data out or as not undoable. | 这个工具的服务器把这一步标为会向外发送数据或无法撤销。 |
|
|
1049
|
+
| `tool_marks_a2a` | ✗ | This sends a request to another agent service, which may act on it. | 这会向另一个智能体服务发送请求,对方可能据此行动。 |
|
|
1050
|
+
| `policy_rra` | ✗ | This deployment's policy asks for your approval here. | 这个部署的策略要求这里由你批准。 |
|
|
1051
|
+
| `artifact_shared_read` | ✗ | This reads something another task shared. (Today the shared-artifact card carries no class word, so this row is not shown.) | 这会读取另一个任务共享出来的内容。 |
|
|
1052
|
+
| `mode` | ✗ | Your current permission mode asks before this step. | 你当前的权限模式要求这一步先问你。 |
|
|
1053
|
+
| `unstated` | ✓ | This tool does not say what it needs permission for, so you are asked. | 这个工具没有说明它需要什么权限,所以要问你。 |
|
|
1054
|
+
| `ordinary` | ✓ | No rule or check singled this step out — steps like it are just not allowed on their own yet. | 没有哪条规则或检查专门针对这一步 —— 只是这类步骤还不能自动放行。 |
|
|
1055
|
+
|
|
1056
|
+
- **Sentences rewritten** (what happened / is anything needed / what to do; old → new; all canonical English — none of these seats holds a run's language):
|
|
1057
|
+
1. `inherit_env.names_withheld` (the person's notice at the first shell call). Old: `The tool shell in this session does not receive N variable(s) of the engine's environment: A — withheld by the credential scrub; the run's inheritEnv ("all", or a list naming them) hands them on; B — the engine's own service credential(s), withheld under every inheritEnv. Values are never shown.` New: `The tool shell in this session does not see N of the engine's environment variable(s). This is a protection: nothing is needed from you[ unless a command needs one of them]. A — held back because it looks like a credential (if a command needs it, name it in the run's inheritEnv, or set inheritEnv to "all"); B — the engine's own service credential(s), never handed on, whatever inheritEnv says. Values are never shown.` Unknown arm: old `The engine could not determine which variables of its environment <shell> does not receive: the execution environment does not say. Variables you set may read as unset there.` → new `The engine could not tell which of its environment variables <shell> can see: the execution environment does not say. Nothing is needed from you unless a command finds a variable unset — then set that variable in the execution environment itself.` The operator's `config.secret_env_scrubbed` notice is unchanged.
|
|
1058
|
+
2. `memory.session_polluted`, the `carry` arm (the default). Old: `Memory for this session is marked externally exposed: <reason>. Ordinary memory entries it writes are still committed … instruction-form files … are withheld and quarantined for host review, and this session's index prose additions are not retained. Contributions this session already committed are challenged …` New: `Memory for this session now counts as exposed to outside content (<reason>). Nothing is needed from you. Ordinary memories this session saves are still kept, each with an engine-written external-origin marker (the content checks run as before); memories that would act as standing instructions (type: feedback, or pinned, triggers or applies-when header lines) are held for the host to review instead of being saved, and this session's additions to the memory index are not kept. What this session saved before now is checked again (if that check cannot run, memory.challenge_sweep_failed says so). Each harvest that holds something back says so as memory.harvest_quarantined.` The `off` arm keeps its pre-provenance bytes (that arm's own promise).
|
|
1059
|
+
3. The shell face's deny refusal (read and write), after its unchanged cause clause. Old read tail: `The read boundary judges the paths this command names — as written and, where a filesystem can be consulted, where they resolve and what a pattern among them matches — and refuses a deny-listed one here rather than escalating it; a file the command reaches without naming it is judged only where a segment deny row (built in, or the deployment's own pattern) reaches into the tree a recursive read walks.` New: `The command was not run, and nothing was read. Paths a command names are checked against this list, as written and where they lead; a file it reaches without naming it, for example inside a folder it searches, is checked only where a deny rule covers that folder. To get past it, the pattern has to be removed where it is set (this deployment's deny patterns, or a deny rule in the permission settings); the agent cannot lift it and no approval does.` Write tail: old `The boundary judges this command's write targets — its output redirections and the operands of … — against the write deny list; a denied write is refused here rather than escalated.` → new `The command was not run. Its output redirections and the files … write are checked against the write deny list. To get past it, …` (same outlet). The two footnotes: old `The working directory at this word is not settled — a move before it may fail or is one this boundary does not follow — so the word was also judged where the command stood before that move (a bare cd stands it in the declared home).` → new `(The command changes directory before this path in a way the engine cannot follow, so the path was also checked from where the command stood before that change; a bare cd counts as the home directory.)`; old `The word was judged as spelled with the environment's declared home: the command may rebind HOME before it, or hands the program a literal ~ it may expand itself — the shell may open another file, and a deny row is refused wherever the command may read it.` → new `(The path was also checked with this environment's home directory standing in for ~ and HOME, because the command may change HOME first, or hand a literal ~ to a program that expands it itself.)` The reach the note states is unchanged (KL-1102-a).
|
|
1060
|
+
4. `config.memory_scope_key` (sync of a `userlocal:` scope), the way out. Old: `Leave this layer out of the sync (isReplicableScope).` New: `Leave this scope out of what you sync; memory meant to travel goes under a scope that does not start with userlocal:.`
|
|
1061
|
+
5. The erase attestation's `status: "complete"` (`MemoryErasureAttestation`): docs only — the attestation is a structured record by contract and gains no sentence; its `status` doc and INTEGRATION-CORE now say `complete` covers the clearance log's held text (rows kept for audit, text erased, `custodyErased: true`; an empty `entryText` there is the erasure).
|
|
1062
|
+
6. The approval card's first line: `headline` above.
|
|
1063
|
+
- **Side check.** A shell call's `description` (the model's purpose line) rides `AskRequest.args` to the approver unchanged (pin H7); a card that does not show it is the surface's.
|
|
1064
|
+
- **Pins.** `test/backlog4130-ask-headline.test.ts` — H1 structural (the clearable classes are a suffix of the class order; every ordered pair folds to the earlier class's own `ruleMayClear`), H2 every class but `content_question` in both locales and that one absent, no engine word, no "a person" / "flagged", H3 real Runner live card canonical + zh-Hans, H4 chokepoint replaces a caller's value / mints none without a class word, H5 durable park row + listing echo canonical + zh-Hans, and a row a canonical resume gives back keeps the parking run's zh-Hans line (a resume does not re-mint it), H6 a damaged stored value reads as absent, H7 `description` reaches the approver, H8 the memory exposure notice; re-pinned to the new sentences: `backlog1963-withheld-env-user-notice`, `backlog3650-child-leg-withheld-notice`, `backlog3200-own-credential-wording`, `backlog939-deny-hit-card-names-pattern`, `backlog1082-shell-read-shadow`, `backlog2970-userlocal-scope`, `backlog994-s1-dual-run` (attribution phrase), `backlog1595-engine-text-catalog` (45 → 74 messages), `prompt-locale-selector` (the composer's default reading), `backlog1320-reader-requirements` (the `headline` read sites), `module-size-ratchet` (seven ceilings, each with its reason).
|
|
1065
|
+
|
|
1066
|
+
### A resume's sentences name what the paused row waits on — a wake is a wake, not an approval — and say how to carry on without engine words (#4124)
|
|
1067
|
+
|
|
1068
|
+
- **Rule.** One noun reading of the parked row (`parkNounOf`, `src/core/park-noun.ts`): approval request / question (a `tool_approval` row whose origin is `content_question`) / plan review / paused run (`resource_limit`) / review / finished task (`task_done`); no row in hand ⇒ "the paused run", never a guessed kind. One composer for a resume cancelled before anything ran (`resumeCancelledSentence`), read by the seven `checkpoint.resume_aborted` rungs; the unstarted-row sentences read the same noun.
|
|
1069
|
+
- **Before.** Seven sentences ended `— the checkpoint stays pending and is resumable with a live signal` (and named `userPromptSubmit`, `resumePreflight`, "before the claim"); the unstarted-row sentences said "the approved action" / "the same approval" for every row, a wake's included.
|
|
1070
|
+
- **Now.** e.g. `the resume was cancelled while this deployment's hook was still checking the wake message: nothing ran, and the finished task is still waiting — wake it again to carry on`; the other rungs: `while this deployment's resume check was still running` · `while the call was being checked again` · `while it waited for the session to be free` · `before it began (the signal it was handed was already cancelled, so it took nothing)` · `while this deployment's usage limits were being read`. Unstarted rows: `<the approved action | the wake | the approved plan | …> never started, so the checkpoint was REOPENED (pending): the same <approval | wake | …> can be resumed again — <resume it again | wake it again>; do not ask for a fresh one`; `declined` / `unknown` say "a fresh <noun>"; `tool_unavailable` says "the tool it needs" (was "the approved tool"). The post-claim abort exit leads `the resume was cancelled before <the approved action | the wake | …> could run — …`.
|
|
1071
|
+
- **Not changed.** Every code (`checkpoint.resume_aborted`, `checkpoint.reopen_failed`, `resume.reopened_unstarted`, `resume.reopen_failed_unstarted`), every outcome and reason word, the row's state; an approval row's unstarted sentences read as before here except the env-failed tail (`— resume it again; do not ask…`) — #4217 (this release, below) then rewrites the `declined` and `unknown` ones.
|
|
1072
|
+
- **Hard-breaking.** None (message text only; consumers branch on the codes). @server @cli @test: a pin on the old wording reds — re-pin on the code.
|
|
1073
|
+
- **Rules.** Spellings of "what this row waits on" 1 (was one hard-coded noun, wrong for four kinds); cancelled-resume sentences 7 literals → 1 composer.
|
|
1074
|
+
- **Pins.** `test/backlog4124-park-noun-sentences.test.ts` — N1 one noun per kind (and none guessed), N2 the composer has no engine word, N3 a wake row's every outcome says the wake and never an approval, N4 real Runner: a wake resume cancelled while its hook checks the message says the finished task is still waiting, wake it again, the row still pending.
|
|
1075
|
+
|
|
1076
|
+
### The sentences a person or the model reads on an approval card, a refusal and the end of a cancelled resume say what they mean — a rule, a hook, a read-only check, a withheld variable, a cancelled resume; two additive members: `detail.remedies` on `inherit_env.names_withheld` and `words` on the two hook-cause entries of `askReasons` (#4196 #4198 #4199 #4207 #4217 #4228; behavior change: sentences only; narrowing: none; widening: none; hard-breaking: none; closed sets: one added, `WithheldEnvRemedyKind`; exports: none added, none removed; wire: two additive members) (@server @cli @client-core @test)
|
|
1077
|
+
|
|
1078
|
+
Every row is old → new, canonical English (none of these seats holds a run's language; no catalog row, no zh-Hans rendering moves). A reader that matched an old sentence reads the new one: branch on the codes and class words, which are unchanged.
|
|
1079
|
+
|
|
1080
|
+
- **#4196 — the card's detail layer (`message`).** (a) A declarative ask rule: `flagged by permission rule "R" (S)` → `the permission rule "R" (S) requires approval before a call like this runs`. (b) A content rule the call could not be read against: `tool "T" needs approval: the call could not be read against the permission rule R (why) — a person must decide` → `… (why), so the engine cannot tell whether that rule forbids it`. (c) The organization lane's same question: `an organization policy rule (R) may reach this call, which could not be read against it (why) — a person must decide` → `… (why), so it needs approval first`. (d) The cross-boundary station's ask: `… executes on a target device — a person must confirm it` → `… — it needs approval before it runs`. The words "a person must …" are gone from card sentences, because the card's reader is the person and the same `message` is also quoted to the model where "you" would be the model. (e) The frame a live steer reaches the model in (the head is unchanged; D-4196-a, ruled C-R1459, open to veto @server @cli @client-core: the reference frames only a message a running turn absorbs and its tail says "rather than as a separate conversation turn", while this engine frames the text when it is taken, before it knows which of the two forms it will meet, so the tail says both): `This is how a message sent mid-turn is surfaced: within the turn that is still running, often alongside the next tool result, rather than as a separate conversation turn. Address the message above as you continue this turn.` → `This is how a message sent while you were working is surfaced: next to the next tool result when you were mid-task, or as the next turn when your last reply was already complete. Address the message above, continuing what you were doing where that still applies.` — the old tail was not true when the run's last reply was already complete and the run opened another turn for the message.
|
|
1081
|
+
- **#4198 — `inherit_env.names_withheld` carries its way out as data.** New member `detail.remedies?: readonly { kind: "inherit_env" | "execution_environment" }[]` — the member family `delegation.ask_unresolvable` already uses for its ways out (a list of `{ kind }` over a closed set, absent when there is none): `inherit_env` (the run's `inheritEnv` hands on the `names` that are **not** `ownCredentials`; the engine's own credentials have no way out, so a notice whose names are all its own carries no `remedies`, and a mixed notice's `inherit_env` is about the rest) and `execution_environment` (`state: "unknown"`: the variable a command finds unset is set in the execution environment itself). The user sentence (rewritten in the previous batch, #4130) is unchanged.
|
|
1082
|
+
- **#4199 — the read-only reader's reasons.** `command "x" is not in the read-only allowlist` → `command "x" is not on the list of read-only commands`; the one verdict tail on 37 sentences (`NOT_AUTO_ALLOWED`, one constant): `— not auto-allowed` → `— so it is not allowed through automatically` (true on the approval card and on a refusal that is not a card: the read-only seat refuses a call with the same words). The option sentence (`option "-o" is not one of the read-only options of "sort"`) is unchanged. About sixty free-text reasons keep their stems (their rewording is a separate design).
|
|
1083
|
+
- **#4207 — a hook cause the hook did not state carries words.** The hook entry of `askReasons` for `verdict_unreadable` / `verdict_unfinished` carries `words` — the engine's own sentence for the cause (`the hook's verdict could not be read` / `the hook did not finish with a verdict`) — so a card that renders an entry's `words` has a line for these two as for any other asker; the request's sentence reads the same entry (one sentence, one mint, `HOOK_CAUSE_WORDS`). **Whose words `words` are is read off `cause`** (the hook's on `asked`; the engine's on these two). The host's account (the answer's `message`) is carried by no entry, as before. A parked row an earlier engine wrote carries the two causes without `words` (absent reads as absent); the row's screen accepts exactly the engine's sentence for the cause and refuses anything else.
|
|
1084
|
+
- **#4217 — the sentences a resume ends on.** `declined`: `the approved action never started, and the store DECLINED to reopen the checkpoint — it is no longer the resolved row this leg claimed (missing, expired, its decided action already disposed, or pending again): read the row before deciding; if it is resolved or gone, a retry needs a fresh approval` → `the approved action never started, and the approval request could not be put back to waiting: it is no longer the one this resume took (it is gone, has expired, was already used, or is waiting again). Check where it stands now before deciding anything: if it is waiting again, resume it again; if it is gone or already used, carrying on needs a fresh approval`. `unknown`: `… and reopening the checkpoint FAILED IN FLIGHT (<why>) — its state UNKNOWN, unprovable from here: re-read it before deciding; do NOT issue a fresh approval on the assumption the old one is dead` → `… and putting the approval request back to waiting failed partway (<why>), so it is not known whether it is waiting now. Check where it stands before deciding anything, and do not ask for a fresh approval until that is known: the old one may still be usable`. Both read the kind of pause (a wake row says "the finished task" and "wake"). The post-claim abort of the parked-resume drive: `the resume was aborted after its row was already claimed — the checkpoint stays consumed and the caller's rollback settles the outcome-unknown terminal (reopening here would strand a pending approval no redemption path can reach)` → `the resume was cancelled after it had already taken the approval request to carry it out: it is no longer waiting, and it is not known whether the approved action began — check what the run left behind before deciding anything (putting it back to waiting here could leave a request nobody can answer)`.
|
|
1085
|
+
- **#4228 ① — the `dontAsk` refusal says what the call needed approval for.** `Permission to use T has been denied because this run is in the dontAsk permission mode (this call needed approval: ordinary). …` → `… (this call needed approval: <the ask's own message>). …` — the class word only when the ask carries no message. The message is display text: one line, capped at 500 characters, the engine's envelope tags defused (the same treatment a hook entry's `words` get). The ask's message says which partition, which reason, which rule (a content rule the call could not be read against; the write gate's generic question reads `approve write to "<path>"?`; the question tool's reads `AskUserQuestion: awaiting a human answer (durable)`); the head, the guidance and the way out (`To let a call like this run: …`) are as they were, and so is the disposition (`deniedBy: "mode"`, `mode`, the class word).
|
|
1086
|
+
- **Not changed.** Every code, class word, origin word, mandate word and decision; no gate, CAS, settlement or resume path reads a sentence. `message` of the three asks above and the two frames keep their members. Old readers of the two new members: zero change (a reader that did not read them does not).
|
|
1087
|
+
- **Hard-breaking.** None. **Narrowing / widening.** None. **Old readers.** A consumer that matched an old sentence literally reads the new one — @server @cli @client-core @test: re-pin on the code or class word (the sentences above are the rows to look up).
|
|
1088
|
+
- **Rules.** Readings of "whose words are an entry's `words`": 1 (read off `cause`); the hook-cause sentence 2 spellings (a literal table in the request's sentence, an absent member) → 1 (`HOOK_CAUSE_WORDS`); the verdict tail 1 constant over 37 sentences (unchanged count, new words); readings of "a notice's way out": the notice's `remedies` joins the one family `delegation.ask_unresolvable` already used (no second name).
|
|
1089
|
+
- **Pins.** `test/backlog4196-card-detail-sentences.test.ts` (C1–C6), `backlog4198-withheld-env-remedies` (R1–R5: scrub only, mixed, own only, unknown, the closed set), `backlog4199-read-only-reason-sentences` (R1–R4), `backlog4207-hook-cause-words` (W1–W4), `backlog4217-resume-end-sentences` (E1–E5, three exits), `backlog4228-dontask-own-question` (D1–D5); re-pinned to the new words: some forty existing test files that carried the old tail, the program-list sentence, the steer frame's tail or the two resume sentences (the commit diff lists them), the #3826 author table (one row for the new `inlineUntrusted` call in `dontAskRefusalMessage`), and the #994 dual-run and program-read-model equivalence pins, which read the two reworded read-only phrases in the recorded words (they hold which verdict, not how it is worded); and the seven pins this branch inherited red from #4130 (the read-station and writer-row sentences, the 1214 carry census and row pair, the 803 request member list, the comment-census word, the phase-API Input floor 1195 → 1198, the catalog identity refreeze).
|
|
1090
|
+
|
|
1091
|
+
### An MCP server that never answers is waited for once per session, not once per message: the session's connection pool remembers a connection attempt that waited out its startup bound, the startup bound nobody declared is 30 s, and `Runner.forgetMcpFailure` lets a host have it tried again (#4187; behavior change; narrowing: two, below, open to veto @server @cli @client-core; widening: none; hard-breaking: none; closed sets: none; exports: two types added, one Runner method added, one constructor option, one member and four optional members added; wire: none — no key added or removed; the user notice `mcp.server_unavailable` gains a sentence and its `detail.sessionId` names the pool's session, the model's `mcp_unavailable` frame gains a "for this session" arm) (@server @cli @client-core @test)
|
|
1092
|
+
|
|
1093
|
+
- **Before.** Every user message is a run, and every run made a new MCP connection pool and dialed every declared server before its first model request. A stdio server that started and never answered its `initialize` was waited for its whole startup bound on EVERY message — 60 s when nothing was declared (the MCP SDK's default), with no frame on the stream while it waited — and the person was told it was skipped on every message. A delegated child or a workflow agent declaring the same server waited again.
|
|
1094
|
+
- **Now — the rule.** A session's MCP pool is held for the session by its session store (keyed on the store object, so every Runner over one store — a pair that differs only in hands — reads one memory); a leg handed no pool takes its session's, and hands it to every leg it spawns (a delegated child, whether or not the root declares a server; a workflow agent, through `WorkflowInternals.mcpPool` and the auto-mounted `Workflow` tool). The pool remembers, per declaration key, a connection ATTEMPT that failed by waiting out its startup bound (`timeout`: the handshake or the first `tools/list` did not answer in time) — the handshake, the first listing and the 401 re-dial a renewed seat answer makes inside it are one attempt; read where the dial begins, written where it ends. That wait is what the memory spares: a key whose last attempt timed out is not dialed again in that session while its pool is kept (a process holds at most 1 024 session pools and lets the least recent idle one go past that, `SessionMcpPools`; a session whose pool was let go dials once more on its next message, the memory having gone with the pool), and the next attempt answers at once with that failure (its wiring row reads `failed` as before). A failure that came back before the bound — connection refused, a stdio server that exits as it starts, an HTTP status, `unauthorized`, a refusal of the engine's own seats or the revocation ledger — costs the next message no wait and is not remembered: the next message dials again, as before. Nor is the leg's own stop (the run was aborted or is tearing down). A live connection under the key comes before the memory. The memory is process-local and not persisted. A key's word is its newest attempt's: an attempt that settles after a newer one has settled, or after a re-dial or a forget, is no word either way (an older failure does not overwrite a newer success, nor an older success clear a newer failure); a forget also drops the dial still in the air under the key, so the next attempt dials instead of joining it; a closed pool recalls and keeps nothing (its caller dials for itself).
|
|
1095
|
+
- **The way back.** A changed declaration is another key (a renewed credential that changes the declaration's content is one). `TaskStream.reconnectMcpServer` (a live run) drops the memory and dials now; its failure is remembered again. A reconnect that waits out the bound is answered to the host — the verb's result and the `mcp.server_redialed` notice say it — and the person gets no `mcp.server_unavailable` for it (7.67.1 told the person once; measured by the release review). Between messages, `Runner.forgetMcpFailure({ sessionId, server })` — on any Runner over the session's store — drops it and answers `{ server, outcome: "forgotten" | "none_remembered" }`; nothing is dialed until the session's next message. A request whose members are not non-empty strings is refused `config.invalid_argument` naming the member.
|
|
1096
|
+
- **The startup bound.** `mcpStartupTimeoutMs` answers the bound in force: the declaration's `startupTimeoutMs`, else `MCP_TIMEOUT`, else 30 000 ms (the reference's dial bound) — never `undefined`. The pool key carries that answer, so two mounts that declare no bound still share a connection.
|
|
1097
|
+
- **What the person hears (`mcp.server_unavailable`).** Once per failed attempt, on the leg that settled it — so once per session per server for a remembered failure, and once per leg for legs that shared one dial (two background children dialing together each tell it). A remembered failure is worded `MCP server "<name>" did not answer within its startup timeout, so it was skipped — its tools are not available to the assistant for the rest of this session, which goes on with the tools of the servers that did connect. It is not tried again unless its configuration changes or it is reconnected; if it is only slow to start, give it a longer startup timeout.` A later leg answered from the memory says nothing. A failure that is not remembered keeps the old per-run sentence, on every message it happens. `detail.sessionId` (and the notice's top-level `sessionId`) is now the session whose pool holds the attempt: for a delegated child or a workflow agent, the session that handed it the pool — the one the person is in — not the child's own. The operator line is unchanged (once per session per text).
|
|
1098
|
+
- **What the model reads (`mcp_unavailable`).** A remembered failure is listed under the reference's `are unavailable for this session:` arm, with no "tried again" clause; any other keeps `are unavailable for now:` and `Each is tried again when the next message arrives; whether it connects then depends on the error shown.`, byte for byte as before. With both kinds, the session list comes first. Whether a failure is remembered is read off the materialization's warnings (`memory`), the one carrier the notice reads too.
|
|
1099
|
+
- **Narrowing (open to veto @server @cli @client-core).** (1) *The default startup bound 60 s → 30 s.* Who is hurt: a server that declares no `startupTimeoutMs`, run where no `MCP_TIMEOUT` is set, and needs 30–60 s to start — its handshake or first listing lands after 30 s (a cold `npx` start on the server host) — it now fails, and — having waited out the bound — is remembered as failed for the session. Compensation: declare `startupTimeoutMs`, or set `MCP_TIMEOUT` (60 000 restores the old bound); the notice says so. (2) *A server that timed out is told once per session, not once per message.* Who is hurt: a surface that counted one notice per message for such a server. Compensation: the notice names the server and the way back; the wiring row still reads `failed` on every leg. A failure that came back before the bound is told on every message, as before.
|
|
1100
|
+
- **Also changed (behavior, not shape).** A server that timed out is no longer retried by the session's next message: one fixed without changing its declaration stays failed in that session until the host reconnects it (`docs/KNOWN-LIMITS.md` KL-4187-b). A failure that came back before the bound is retried by the next message, as before. A child's or workflow agent's notice is attributed to the session whose pool it leased from. Two children spawned in one turn under a run that declares no MCP server of its own now lease one connection of a server their definition declares (each dialed its own before: there was no run-tree pool without a declaration — `docs/KNOWN-LIMITS.md` KL-2595-a's second arm is retired). A library caller that hands `materializeMcpTools` its own `McpConnectionPool` gets the same memory on that pool. `ToolExecuteContext.mcpPool` (new, optional) carries the run tree's pool; `MountedMcp` (not exported) no longer carries it.
|
|
1101
|
+
- **Known limits.** The session's FIRST message still waits the startup bound of a server that never answers, with no frame saying so (KL-4187-a); a connection is still released with its run (a healthy stdio server is started once per message, as before). A message joins a dial still in the air from an earlier message's background child (KL-4187-c); a session released from its store and reused under the same id keeps the memory (KL-4187-d); a declaration changed and changed back finds its old failure (KL-4187-e); a failure that becomes remembered under the same error text (a forget or reconnect landed while its earlier attempt was in the air) is told to the person as the session's but the model keeps the earlier "for now" frame, because what the model was told is recorded by server name and error text (KL-4187-f; slice C re-cuts that frame).
|
|
1102
|
+
- **Divergences from the reference (CC 2.1.321), each open to veto.** D-4155-a: the first dial is the session's first message that needs the server, not session start. D-4155-b: that first message waits the bound before its first request (the reference does not block). D-4155-c: the person is told once per session per server for a failure that timed out (the reference says nothing in the session; its status panel shows it). D-4155-d: a remote server that timed out is remembered like a stdio one (the reference retries a remote server three times in the background); a remote failure that came back before the bound is retried by the next message. D-4155-h: between messages, recovery is "forget, the next message dials", not "dial now". D-4155-i: a failure that came back before the bound — `unauthorized` included, as the reference's needs-auth is not a failure — is not remembered.
|
|
1103
|
+
- **Rules.** Replaced: the pool's lifetime (run tree → the session's, held by the store), what it keeps of a failed dial (nothing → an attempt that waited out its bound), the default bound (the SDK's → the reader's 30 s), the child hand-off predicate ("the parent declared MCP and the child declares a server" → "the run has a pool"), the model sentence (one arm → the arm each failure's memory says), the person's sentence, the notice's session (the leg's → the pool's). One rule says what is kept — "a failure that made the person wait the bound" — where two exceptions (`unsent`, `unauthorized`) were listed. Added: the session pools seat, the recovery verb. Retired: "every run re-dials every declared server".
|
|
1104
|
+
- **Not breaking at the type level — `mcpStartupTimeoutMs` answers `number` (a wire-manifest reading).** It answered `number | undefined` and now always answers the bound in force (the declaration's, else `MCP_TIMEOUT`, else 30 000 ms): every caller still compiles, and a replacement typed as `typeof mcpStartupTimeoutMs` that answers `undefined` is the one that no longer does. A caller that read `undefined` as "nothing declared, use my own default" now reads 30 000 there (the startup-bound bullet above).
|
|
1105
|
+
- **Exports.** Added: `ForgetMcpFailureRequest`, `ForgetMcpFailureResult` (types); `Runner.forgetMcpFailure`; `McpConnectionPool.attempt` / `forget` / `idle`, its constructor option `{ sessionId }` and member `sessionId` (the session a pool is held for); optional members `RunInternals.sessionMcpPools` (the store's pools, which the Runner fills), `WorkflowInternals.mcpPool`, `RunWorkflowToolDeps.parentMcpPool`, `ToolExecuteContext.mcpPool` (the run tree's pool, handed down). Removed: none. `mcpStartupTimeoutMs` now returns `number`.
|
|
1106
|
+
- **Pins.** `test/backlog4155-mcp-session-pool.test.ts` (real Runner, real stdio children, real loopback HTTP): ① a silent server's second message neither waits nor dials · ②a a changed declaration dials · ②b another session dials · ②c another store dials · ②d two Runners over one store share one memory and one forget · ③a forget ⇒ the next message dials, a server fixed meanwhile connects · ③b a forgotten server that fails again is remembered again · ③c the request's members refused by name · ③d the run-level re-dial dials now and is remembered · ③e a forget while the attempt is in the air: that failure is told for its run and the next message dials · ④ a healthy server connects on both messages with the same tools · ⑤a a first listing that never answers is remembered · ⑤b a 401 cured inside the attempt leaves no memory · ⑤c a 401 that stays is asked again each message, neither waits · ⑤d the ledger's refusal is not remembered · ⑤e the run's own abort is not remembered · ⑥a the notice once per session with the way back · ⑥b the model's "for this session" arm once per session · ⑥c a refusal not remembered keeps the per-run notice · ⑦a / ⑦b a delegated child (also under a root that declares nothing) dials nothing · ⑦c / ⑦d a workflow agent (host-started; launched by the auto-mounted `Workflow` tool) dials nothing · ⑧ the reader answers 30 000 and the key carries it · ⑨ a live connection comes before the memory · ⑩ (pool) an older settle is no word either way, a forget drops the dial in the air, a closed pool recalls nothing, a detached connection's lease keeps the pool from reading idle · ⑪a–c a 503, a refused connection and a stdio server that exits at start are not remembered (the next message dials, and connects once fixed) · ⑪d a first initialize that answers after the bound is (the next message does not dial; the notice says to lengthen the bound) · ⑫ a child's attempt is told once, on the root's session · ⑬ the next message joins a background child's dial still in the air (KL-4187-c, pinned as it stands) · ⑭ a failure that becomes remembered under the same error text is not told to the model again (KL-4187-f, a residual pinned as it stands); ③e and ⑤d read the model's "for now" arm. `turn-attachments-g1-tools-delta` adds the remembered arm and the two-kind order. Re-pinned: `backlog1552` S2d, `backlog2060` E-2084.b, `mcp-env-knob-hygiene` (the `MCP_TIMEOUT` fallback), `backlog2841` ⑫ (it leases under the key the dial uses, which now carries the reader's 30 000), `backlog2595` A7b (one connection for the two siblings, above), `design393-s7-runner-methods` (`forgetMcpFailure` on the prototype), `backlog3186-store-options-gate` (`McpConnectionPool`'s bag is not a store); `flaky-guard` declares the pin file's wall-clock rows; the module-size ceilings of five files and the phase-input floor (1195 → 1198) are raised with their reasons.
|
|
1107
|
+
- **Consumers.** @server: the `/mcp/reconnect` 409 sentences that say "a resumed continuation takes a fresh first listing" / "the next submit … materializes every declared MCP server from scratch" are no longer true; `forgetMcpFailure` is the between-messages verb (on either of the pair of Runners over the shared store); the `/mcp` panel dials without the session's memory, so it can read `connected` while the session remembers `failed` (KL-4187-b). @cli @client-core: `mcp.server_unavailable` for a server that timed out arrives once per session per server with the new words; a child's arrives on the root's session id; a failure that came back before the bound arrives on every message, as before. @test: the session's second and third messages to a server that never answers start their first model request without the 30 s wait; a server answering 503, refusing the connection or exiting at start is dialed again by the next message.
|
|
1108
|
+
- **Coordination.** A sentence elsewhere in this window's work saying every run still dials a server is no longer true once this lands; the later of the two to merge restates it.
|
|
1109
|
+
|
|
1110
|
+
### What the engine says to an operator leaves by one door — every write the build made to the process's stderr itself (`console.warn`, `process.emitWarning`, the opt-in debug `console.error` traces) is an `EngineNotice` through `deliverEngineNotice`: the seat in hand, else the process's default seat, else one line from the throat's guarded writer, after the same-principal secret-shape pass (#3832 = #3821 S1; ruled C-R1276, C-R1412; behavior change: a host that installed a notice seat receives these facts there instead of on stderr; narrowing: two, below, open to veto @server; widening: none; hard-breaking: one closed-set rename, `config.forward_sink_failed` → `config.host_sink_failed`, open to veto @server @client-core @cli; closed sets: `EngineNoticeCode` +14 −1 from this section, the release's count below; exports: none; wire: notice codes) (@server @cli @client-core @dgx @test)
|
|
1111
|
+
|
|
1112
|
+
- **Rule.** The engine's own words to an operator have one exit, `deliverEngineNotice`: the seat in hand (`RunnerDeps.onNotice`, a band's or an environment's `onNotice`), else the process's default seat (`setDefaultEngineNoticeSeat`), else one process line. That line is written by the one writer in the build (`writeProcessLine` in `src/core/engine-notice.ts`), guarded: a throwing console never fails the call that spoke. Places that hold no seat (a store, a tool band, a process-level knob, a memory engine a host built, the durable-agent lane, a policy built outside a run) hand `undefined` and reach the default seat. Every sentence is unchanged; only its channel moved.
|
|
1113
|
+
- **Before.** 32 writes in 20 build files wrote to the process's fd 2 themselves (31 on this branch's base, and the stage's parked-card reader brought a 32nd) — 22 `console.warn`, 4 `process.emitWarning`, 6 debug `console.error` — and the throat's own absent-seat arm wrote unguarded. Apart from the throat's two, none reached a host's notice seat or the throat's text pass, and each carried its own posture toward a throwing console (a `try`, a safe-notifier `onError`, or none).
|
|
1114
|
+
- **Now — the codes.** `config.host_sink_failed` (a run-internals sink, the delegation-lifecycle observer or the subagent display sink threw or rejected and was contained; `detail: { site: "run_internals" | "delegation_lifecycle" | "forward", cause }`); `config.seat_malformed` (`onDelegationLifecycle` or `onSecretEnvScrub` present but not a function; `detail: { seat }`); `config.env_knob_adjusted` (`detail: { name, applied? }`); `config.read_deny_home` (`detail.kind`: `bad_value` / `unanswered` / `auxiliary` / `config_root` / `placement_lost`); `config.permission_rules_skipped` (`detail: { count, codes }`); `config.process_start_time_unavailable` (`detail: { reason }`); `store.tail_segment_unjudged` (`detail: { path }`); `store.lock_retained` (`detail: { lock: "directory_fence" | "boot" | "mailbox", path, pid }`); `memory.notice_unscoped` (no incident sink; `detail: { stage }`); `delegation.settlement_external` (`detail: { sessionId }`); `delegation.settlement_write_failed` (on the mounting runner's seat; `detail: { verdict, sessionId, fork }`); `delegation.durable_row_unpersisted` (`detail.kind`: `terminal_unrecoverable` / `poisoned` / `admission`, `detail.handle`); `delegation.parked_card_unreadable` (the store sweep cannot read a parked agent's approval card, so its row stays parked: `detail: { handle, scope, cause: "decode" | "no_state" }`, once per (scope, handle) per process — the reader that said it as a process warning is #3983's, `parkedCardState` in `src/core/background-agent-store.ts`); `engine.debug_trace` (`SEMA_DEBUG_SHUTDOWN=1` / `SEMA_DEBUG_WORKFLOW_JOURNAL=1`; `detail.topic`). All audience `operator`, minter `engine`.
|
|
1115
|
+
- **Hard-breaking (open to veto @server @client-core @cli).** What broke: `config.forward_sink_failed` is gone from `ENGINE_NOTICE_CODES`, `NOTICE_AUDIENCE` and `NOTICE_MINTER`, and the display sink's failure is now `config.host_sink_failed` with `detail.site: "forward"` (the old code's `detail.site` read `"runtask.onForwardEvent"`); the sentence is unchanged. Who is affected: a consumer that mirrors the notice catalog — server's code-catalog pins, client-core's generated notice tables, cli's wire-consumption manifest; none of the three routes or filters on the code (read on each repo's origin/main). How to migrate: regenerate or re-pin the mirrored rows to `config.host_sink_failed`, and read `detail.site === "forward"` where the old code was matched.
|
|
1116
|
+
- **Closed sets — the release's count of the notice catalog.** `ENGINE_NOTICE_CODES` 135 (7.67.1) → 153 (7.68.0), +19 −1, counted from the array in `src/core/governance-codes.ts` in both trees. Added — fourteen in this section: `config.host_sink_failed`, `config.seat_malformed`, `config.env_knob_adjusted`, `config.read_deny_home`, `config.permission_rules_skipped`, `config.process_start_time_unavailable`, `delegation.settlement_external`, `delegation.settlement_write_failed`, `delegation.durable_row_unpersisted`, `delegation.parked_card_unreadable`, `memory.notice_unscoped`, `store.tail_segment_unjudged`, `store.lock_retained`, `engine.debug_trace`; one each in another section: `route.credential_mismatch` (#3002), `permission.allow_rule_not_used_in_auto` (#1367), `task.undelivered_input_dropped` (#3954), `instructions.family_undeclared` (#4179), `task.completed_past_limit` (#4184). Removed: `config.forward_sink_failed` (this section). Each other section says `+1`; this is the one total.
|
|
1117
|
+
- **Behavior changes (disclosed).** A host with a default seat or an `onNotice` now hears these facts there and no longer sees them on stderr. The three `sema durable-agents:` lines are no longer process `warning` events — a `process.on("warning")` listener hears nothing; they are `delegation.durable_row_unpersisted`. The opt-in debug traces leave by the writer (`console.warn`, still fd 2) or the default seat, no longer `console.error`.
|
|
1118
|
+
- **Narrowing — open to veto (@server).** (1) A host with no notice seat whose console throws: `rebindCredentialOrigin` no longer throws there — the binding moves and that one line is lost (it used to throw and leave the binding where it was). Who: only a host with no seat and a throwing console; with a seat the rebind is told exactly once before the binding moves, as before. Remedy: install a notice seat (C-R1412). No other caller relied on the throat's absent arm throwing (searched: callers that catch it, and comments that say they rely on it). (2) A seat is never handed a notice its own delivery raised — synchronously, through a second seat, or from a timer or promise it started — that notice is printed as a process line instead. Who: a forwarder that raises engine notices while handling one (it sees those as process lines, not as calls of itself); remedy: none needed for a forwarder that only logs and routes.
|
|
1119
|
+
- **Rules.** Writers of the process's fd 1 / 2 in the build: 32 sites → 1 (`writeProcessLine`); postures toward a throwing console: one per site → one, in the writer (three `try` wrappers around the throat removed); the host-seat failure account: two codes → one, minted in one place (`hostSinkNotifier`, `src/core/runner/forward-sink.ts`); child output streams given as descriptors: two spawn sites → one function (`shellChildStdio`, the shell child's spool pair). Added: one re-entry guard (an `AsyncLocalStorage` of the seats being delivered to).
|
|
1120
|
+
- **Gate.** `test/backlog3821-process-write-single-source.test.ts` reads the files `tsconfig.build.json` compiles off the syntax tree: outside `writeProcessLine` no value reference to `console` (`globalThis["console"]` included), no `process.stdout` / `stderr` / `emitWarning` (any spelling or import), no write to fd 1 / 2, no child `stdio` that inherits or cannot be read (index 1 / 2 only `"pipe"` / `"ignore"` / `"overlapped"`; a non-literal whole value reds), no `execFileSync` / `execSync` without `stdio`; `shellChildStdio` is the one place a child's streams are descriptors, and it answers the child's standard input too (a pipe when the start carries `stdinFrom`, closed otherwise — the exec site passes the one flag and spells no stdio of its own; `test/backlog3821-shell-child-stdin.test.ts`); `StdioClientTransport` is built only in `buildTransport` (the MCP child's stderr pipe is #3833, the next slice).
|
|
1121
|
+
- **Pins.** `test/backlog3832-notice-throat.test.ts` (the writer after the text pass; a throwing console on both arms; re-entry: a seat raising `config.permission_rules_skipped` and `config.read_deny_home` from its own work, two seats raising for each other, a timer started by the seat; the guard is per delivery; one real-path cell per former site kind). Re-pinned: `backlog3983-reconcile-unreadable-card-body` G3a / G3b / G3c / G4b (the announcement is read on the default notice seat, the same sentence and the same count, with the `cause` word; G3d the `no_state` cause); `backlog802-forward-sink-containment` (code and `detail.site`), `backlog1844-credential-rebind` R15 (+ R15b, the seat told once before the move), `issue22-crossprocess-org-admission` / `agent-park-jian3a-guarded-cas` / `agent-park-reviewf1-rollback-honesty` (the default seat instead of `process.on("warning")`), `env-knob-gapfill` / `reap-liveness-gate` / `task-registry` (the traces on `console.warn`), `backlog1896-announcement-scope` (+ `memory.notice_unscoped`).
|
|
1122
|
+
- **Known limit (KL-3821-c).** The re-entry guard follows the async context a delivery starts; a seat that hands its work to a queue or worker built outside the delivery is not covered there, and a seat that never settles keeps its own backlog outside the engine's bounds. The other side: long-lived work a seat starts inside a delivery (a timer, a connection, a worker it builds lazily) carries the guard for its whole life, so a later, unrelated notice raised there is a process line, not a call of that seat.
|
|
1123
|
+
- **Consumers.** @server: the core-side writes now arrive on the two seats it installed (zero routing code); re-pin tests that read core's raw stderr or `process.on("warning")`; the three catalog pins for the renamed code; a forwarder that raises engine notices while handling one now sees those as process lines. @dgx: an embedder with no default seat still gets the lines on fd 2, now through the secret-shape pass; installing `setDefaultEngineNoticeSeat` routes them to its logger. @client-core @cli: the renamed catalog row. @test: black box — a host with a default seat sees a file store's retained boot lock, a skipped permission rule and an adjusted env knob as notices and nothing on stderr.
|
|
1124
|
+
|
|
1125
|
+
### A file restore that fails the task or the verb says its per-file ledger as data — `rewind.restore_failed` carries `refused`, `failed` and three counts on its `detail`, from the one mint both of its restore throw sites share; the refusal of a history with no address (a session whose row cannot be read) mints the same code with no `detail` (KL-4205-a) (#4205; client-core request; behavior change: one coded error and its failed terminal's coded account gain members, and the terminal's coded snapshot drops an array a bound cut anything inside whole; narrowing: one, below, open to veto @server @cli @client-core; widening: one, below; hard-breaking: none; closed sets: none; exports: none; wire: additive — new members of `detail`, below — and a `terminal.detail` array read cut under a bound is now absent whole) (@client-core @server @cli @test)
|
|
1126
|
+
|
|
1127
|
+
- **Before.** The ledger of a restore that failed the task — which files were refused and why, which failed, how many had already been written — existed only inside the error's sentence. The code was also thrown in two places that did not agree: the store's operation-level failure carried `detail = { code, permanence }`, and the run's default refusal of a PARTIAL restore (no `acceptPartialRestore`) threw the same code with the same ledger in prose and no `detail` at all. A host that wanted "which files, and how many were already changed" parsed a sentence, and on the partial arm — the one where files had been changed — it had no coded account to read.
|
|
1128
|
+
- **Now.** One mint, `restoreFailedError` (`src/core/runner/file-history-restore.ts`), builds the error at both throw sites. Its `detail` is the store's judgment when the store stated one (`code`, `permanence`, as before) plus the restore's ledger in the success arm's own row types: `refused` (`{ path, reason, class }`) and `failed` (`{ path, reason }`) as the store answered them, and three counts — `filesChangedCount` (the sentence's "applied N"), `refusedCount`, `failedCount`. The members are always present on what this mint builds (empty lists, zero counts when the store listed nothing), on the thrown error of `Runner.rewindFiles` and on a failed run's `terminal.detail` (`{ kind: "coded", detail }`) alike. A second mint of the same code is not this one: `unaddressedRewindRefusal` (`src/core/runner/prepare-file-history.ts`; the run seat and `Runner.rewindFiles`, for a session whose row cannot be read so its history has no address) throws `rewind.restore_failed` with no `detail` — nothing was restored and nothing is listed (KL-4205-a, #4275). The terminal's coded snapshot is bounded (64 entries per container, 4096 characters per string, depth 4, 256 values and 16384 characters in all): the counts are declared before the lists, so the snapshot admits them first, and a list rides only WHOLE — past a bound it is absent there (about 62 refused rows is the value bound), never a cut list — which is why the counts ride beside the lists: absent is not empty. The thrown error is not bounded.
|
|
1129
|
+
- **Narrowing (open to veto @server @cli @client-core) — the coded snapshot's array law, made true on every bound (correction; no new key, no new word).** The snapshot's written law is "an array is one fact" — an element left out drops the array — but it held only for an element that was itself left out. A record element a bound cut a MEMBER of (a row with no `reason` or `class` once the 256-value or the 16384-character budget ran out, a member over the 4096-character string bound, a member nested past depth 4, a key that did not fit the characters left) stayed as a cut row, and a dropped array kept the budget it had spent, so the members after it were squeezed out of the account. Now an array any bound cut anything inside — an element, or a member of an element record, a nested array dropped included — is absent whole. A record that is not inside an array is read as before (admitted member by member as far as the budget goes). **What narrows (measured, base against head, real Runner and real stores):** an array that used to survive as a cut row is now absent — 64 three-member rows where the 256-value bound falls inside the last row, a row holding a nested array over the 64-entry bound, a circular row, a row with one member over the 4096-character string bound (the other rows of the list go with it: the whole list is absent, not just that row), and the same one level down. **Who is reached:** a host-supplied store, hook or brain that throws a coded error whose `detail` holds an ARRAY OF RECORDS (a list of rows with members) near a bound: its `terminal.detail` read the rows it could fit, cut, and now reads no list. An array of scalars is read as before unless a bound cuts it; a list the producer states a count beside (as `rewind.restore_failed` does) tells absent from empty. Producers in `src` (measured): `grep -rnE '\b[a-z]+\.detail = ' src` finds 32 throw-site assignments; one builds an array of records, `policyNameRefusal` (at most 32 entries of at most 128 characters, sized to these bounds: 195 values in the worst case and about 7K characters), so no producer found by that grep changes its account. That grep form does not see a producer that sets `detail` through a constructor parameter or `Object.assign` and was not extended to them; the ones read carry scalars or scalar arrays (inferred: no `src` account changes). A coded detail that overflowed a bound used to lose its last row's members and the members after it; now its array is whole or absent. Pins: `test/backlog1399-terminal-detail-projection.test.ts` C4d (each bound just inside and just past, the dropped array's budget, the nested array, the record control).
|
|
1130
|
+
- **Widening (the same correction; @server @cli @client-core).** An array a bound drops now gives back what it had spent, so the members after it are judged as if it were not there: a sibling record after a dropped array, and siblings that exactly fill the 256-value budget (the next member is then cut, as before), now ride where the dropped array's spend used to squeeze them out (measured). A detail that read short of members after an oversized list now reads them.
|
|
1131
|
+
- **Not changed.** The sentence, the code, the success arm (`complete` / `partial` answers are returned exactly as the store gave them, no member of the failure's on them), the tolerated arm (`acceptPartialRestore`: the `restore_partial` note keeps `{ code, message }` and its ledger in the sentence), `rewind_snapshot.unresolvable` (an absence, no detail), what is restored or refused.
|
|
1132
|
+
- **Side effect, additive.** A partial refusal whose store also stated an error now carries that error's `code` / `permanence` on `detail` as the failed arm always did (the one mint reads the same member for both).
|
|
1133
|
+
- **Rules.** Throw sites of `rewind.restore_failed` that build their own error: 2 → 1. Readings of the ledger's counts (sentence, members): the sentence's and the members' come off the same lists in the same function. Special cases: 0.
|
|
1134
|
+
- **Wire.** Additive: `detail` of `rewind.restore_failed` (the thrown error's own member, and `terminal.detail.detail`) gains `refused`, `failed`, `filesChangedCount`, `refusedCount`, `failedCount`; `detail` is an open record (`CodedFailureDetail`), so no declared type or closed set moves. The array law (Narrowing, Widening above) changes what an existing `terminal.detail` member can hold — an array is whole or absent, never a cut list — with no change to a declared type.
|
|
1135
|
+
- **Pins.** `test/backlog4205-restore-failed-ledger.test.ts` — the run seat's default partial refusal on the real file store (members element-wise equal to the sentence's numbers, a link-refused row named), a torn boundary through the verb and a run (the same detail both ways, nothing moved), a ledger inside the terminal's bound (rides whole) and past it at the value bound (62 rows whole; 63, 64, 70 absent), at the character bound (long paths) and at the string bound (list absent whole, never a cut row, counts equal the sentence's; the verb's error still whole), the one mint's copy of the store's answer and its member order (counts before lists), the success arm's key set, the tolerated arm's note keys; `test/backlog1468-t2-file-history-permanence.test.ts` re-targeted to the new exact detail (two cells; the store's judgment members unchanged).
|
|
1136
|
+
- **Consumers.** @client-core: read `filesChangedCount` / `refusedCount` / `failedCount` whenever `detail` is there and the lists when present — a `rewind.restore_failed` with no `detail` is the no-address refusal above (KL-4205-a), not a ledger of zeros; a list absent beside a count above zero means the ledger did not fit the terminal's bound, and the sentence still names every row. @server: carry `detail` through the way it carries `permanence` today (the code's HTTP face is yours). @cli: none. @test: the black-box read is a failed rewind on a real tree with one path made a link.
|
|
1137
|
+
|
|
1138
|
+
### The parked row's `riskDescriptor` carries the live card's two positions — `riskAxes` (`irreversible?`, `egress?`, three-valued), one type and one mint with `AskRequest.riskAxes` (#4206; server and client-core requests; behavior change: one optional member on the descriptor a park mints; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none (one type named in place); wire: additive — `RiskDescriptor.riskAxes?`, below) (@server @client-core @cli @test)
|
|
1139
|
+
|
|
1140
|
+
- **Before.** The descriptor on a parked approval gate said which safety axes TRIPPED the gate (`axes`) and nothing about what the call is, so a host reading the parked row, an inbox or a queue could not answer "can this be undone / does it send data out" — the live card could (`AskRequest.riskAxes`). Two faces, two answers to one question.
|
|
1141
|
+
- **Now.** `RiskDescriptor.riskAxes?: CallRiskAxes` — the same member name and the same type as the card's — read for the same call from the same projection the card is minted from (`riskAxesOf`, the policy chain's, handed to the park the way its rule-offer and Skill-binding factories are). `buildRiskDescriptor` takes `riskAxes?` as an input and copies it (each member only when it is a boolean; none left ⇒ no member): it judges nothing about the call's axes. Each member is three-valued: `true` and `false` are judgments, absent is "not judged", and the whole member is absent when neither was judged — and on every row parked before this member existed, which a reader reads as the same "not judged". `axes` is untouched: it is the tripped set (a shell-seat ask trips `shell`, never `irreversible`), `riskAxes` is what the call is — two questions, said in the member's doc and in INTEGRATION-CORE.
|
|
1142
|
+
- **Not changed.** `axes`, `severity` (still derived from `axes`), every other descriptor member, the card's `riskAxes` (its inline type is now the named `CallRiskAxes`, same members), the gate kind, what parks, and any decision that reads the descriptor (none does: display / triage metadata).
|
|
1143
|
+
- **Residual (named; #4235 closes it).** The row's `riskAxes` is a SECOND reading, not the card's own value: the card and the park each call `riskAxesOf` for the call, at different moments (the park mints at the gate, the card at the resolution; they are alternatives except when an approver answers `unavailable` and the ask is routed to the park). The projection reads the call's tool seats, not its arguments, so the two agree unless the roster's seats move between the two reads — a host's MCP re-dial while the card is up, then an `unavailable` answer. Then the row can say what the card did not (measured: a card with no `riskAxes`, a row with `{ irreversible: true, egress: true }`), while the same row's `axes` and `severity` still come from the gate's reading taken before the re-dial (`safety` is read once at the gate; `riskAxes` is not). Read the card for what the approver was shown and the row for what the call was when it parked (@server @cli @client-core). A single reading shared by the card, the park and the three inherited-lane card mints is #4235.
|
|
1144
|
+
- **Rules.** Functions that mint the call's `riskAxes`: 1 → 1; readings of it for one call that parks after a card: two (the card's, the park's — see the residual); types spelling the two positions: 2 inline → 1 named. Special cases: 0.
|
|
1145
|
+
- **Wire.** Additive: `RiskDescriptor.riskAxes?` (`CallRiskAxes`: `irreversible?`, `egress?`); `buildRiskDescriptor`'s input gains the optional `riskAxes`.
|
|
1146
|
+
- **Pins.** `test/backlog4206-park-risk-axes.test.ts` — five cells (egress, irreversible, both, neither judged, a never-irreversible read tool) each run the same call through a real Runner with a live approver and unattended with a checkpoint store, row's `riskAxes` deep-equal to the card's (absent versus absent included) and `axes` as before; a shell-gated call (`axes: { shell }`, the card's `riskAxes`); the `unavailable` re-route arm (equal when nothing moved); the residual, named — an MCP re-dial while the card is up then `unavailable`: the card carries no `riskAxes` and the row carries both axes while its `axes` stay `{}` (the cell's title says it is the residual #4235 closes, and goes red the day the reading is shared); the builder's four unit cells (copies, never derives from `safety`; non-boolean members dropped; a hostile value — an accessor member, a revoked proxy — neither throws nor varies it; the descriptor owns its copy).
|
|
1147
|
+
- **Consumers.** @server: project `riskDescriptor.riskAxes` on the parked row, inbox and queue under the card's own key; a row without it reads "not judged" (no migration). @client-core: the same two positions on the parked card as on the live one; what the approver was shown is the live card's, what the call was when it parked is the row's, and they can differ after a re-dial (the residual). @cli: the same reading rule where a parked row or a card is shown. @test: park the same call attended and unattended and compare the two answers; with an MCP re-dial between the card and an `unavailable` answer the row may carry axes the card did not.
|
|
1148
|
+
|
|
1149
|
+
### A park is minted only on a leg someone resumes — one leg fact with two sources (the lane, the host's `durableApproval: false`); a synchronous child parks nothing and its refusal names the way out (#4111 = #4087 / #4091 S1; ruled C-R1405 with its notes and C-R1413; behavior change: on a leg nobody resumes an ask is refused with the way out instead of parked; narrowing: two, below, open to veto @server @cli @client-core; widening: one, below, open to veto @server; hard-breaking: three, below — `TaskSpec.durableApproval` and `Checkpoint.durableApproval` gain `false`, the Workflow mount's `RunWorkflowToolDeps.parentDurableApproval` is renamed, a background child on a lane that cannot park ends `completed` with `refusedAsks` — open to veto @server @cli @client-core @sdk @test; closed sets: `ParkLaneReason`, `SuspendLaneReason` and `CheckpointErrorReason` each +`leg_not_resumable`, new `ApprovalParkNotResumable` (`lane` / `host`), `REVIVED_ROW_CLEARED_FIELDS` +`refusedAsks`; exports: +`AskUnresolvableCause`, `RefusedAsk`, `DurableApprovalOptIn` (types), none removed; wire: `TaskSpec.durableApproval` widened, `TaskResult.refusedAsks?` and its row and frame twins, additive)
|
|
1150
|
+
|
|
1151
|
+
hard-breaking: `TaskSpec.durableApproval` gains a third state `false` (type `DurableApprovalOptIn | false`); `Checkpoint.durableApproval` the same; closed sets `ParkLaneReason` / `SuspendLaneReason` / `CheckpointErrorReason` each +`leg_not_resumable` (the third, and the earlier release's synchronous-child rows that now read it: C-R1405 note 1); a synchronous child / fork / background child on a lane that cannot park mints no decision row any more, and the synchronous-child block (`parentSessionId` / `parentTaskId` / `rootSessionId` / `rootSessionIncarnation`) is no longer written on any new row; `RunWorkflowToolDeps.parentDurableApproval` (the host's opt-in object for workflow children) is renamed `parentDurableApprovalForChildren` — the old name is now the new clamp row's retired per-seat dep and is refused `config.deps_retired_key`.
|
|
1152
|
+
|
|
1153
|
+
- **Rule.** A park is minted only on a leg someone can resume. One leg fact, `approvalParkNotResumable` (`approvalParkNotResumableOf`, derived once by the driver), two sources: `lane` — this run started the leg on an engine lane that holds no park of it (a synchronous child or fork, a background child whose lane cannot park, an observer sidecar, a retained agent's wake cycle; a background lane that can park and a workflow whose park lane is armed say they hold the leg's parks on its internals, `RunInternals.parkLaneHeld`) — and `host` — `TaskSpec.durableApproval: false`. Where the fact holds, every decision park (a permission ask, the safety vocabulary's included; a question; a plan review) takes the store-less reading, and a `lane` leg mints no resource park either (`suspendLane.reasons` names `leg_not_resumable`); under `forceDurableGate` such a leg says `config.durable_gate_unavailable` / `leg_not_resumable` and refuses its asks unconsulted (#4079, unchanged). The entitlement synthesizes an opt-in on every leg whose parks can be resumed.
|
|
1154
|
+
- **Before (v7.67.0).** The fact had one source, read as a proxy — "a delegated leg its parent forwarded no opt-in to" — so the lane's real question (does anyone resume this leg?) was answered by whether an opt-in was forwarded, and three decision arms never read it: the safety vocabulary's park, the plan review, the plan pair. A synchronous child parked `irreversible_ask` / `plan_review` rows nobody could resume (decided on the same Runner with the child's own tools at hand, the tool ran and the result went to the resumer alone — the parent never heard; decided anywhere else, in a deployment with a tool policy, `resume.parent_constraint_missing` and the row stayed pending forever; with no policy, a fresh runner decided it and the tool ran once, the parent again never hearing), and its parent read `status: suspended`. Under a forced root, a background grandchild on a lane that CAN park behind a synchronous child was refused everything.
|
|
1155
|
+
- **Now.** A synchronous child's unattended ask is refused on the spot and both readers are told what to do next: the child's tool text (the deny exit's remedy) and the person's `delegation.ask_unresolvable` notice name the way out — have a person who is present approve it live, or delegate the work as a background agent on a lane that can park (its approvals wait and its result comes back). With a live approver or a live question face on the run nothing changes: the ask is put to them and answered in-stream. A plan-review request on such a leg is refused to the call that made it: that call's result is an error that says the plan was presented to no one and names the way out — first "give the plan in your final report" (whoever reads the leg's result can present it; CC 2.1.321's ExitPlanMode answers a sub-agent the same way), then the shared roads — never the tool's own "presented" text, and the request leaves its `review.dropped` trace. `durableApproval: false` is accepted by the task door (it was refused `config.invalid_argument`): decision parks are not minted, the plan pair is mounted only as on a store-less leg (`EnterPlanMode` never, `ExitPlanMode` only under `interactiveTools: true`, answered live), resource parks are unchanged and record `false`, a resume whose config states no `durableApproval` keeps it, and every child carries it (`CLAMP_SET.durableApproval`, `ToolExecuteContext.durableApproval`) — a forwarded opt-in never lifts it, a revived child born under it included. A row a synchronous child parked under an earlier release is listed `rowVerdict` `damaged` (`checkpoint.unsupported_version`, reason `leg_not_resumable`, a fixed sentence with the way out) and refused at the decision's first rung with the same sentence, nothing consumed.
|
|
1156
|
+
- **Not changed.** A root leg (store + opt-in, or the entitlement's synthesis) parks as before; a background child on a lane that can park parks as before (also with no forwarded opt-in for an irreversible- or egress-marked tool's ask — its lane holds the park; a plain policy ask parks there only under an opt-in, as on a root); a host resuming a delegated leg's row (`Runner.resume`, no internals) reads only the host's word; workflow agents of an armed workflow park as before; a store-less leg reads as before.
|
|
1157
|
+
- **Hard-breaking — what changed, who is hurt, what to do (open to veto @server @cli @client-core @sdk).** (1) What changed: in an unattended deployment a synchronous child that meets an approval is refused on the spot with the way out, where it used to mint a card; `TaskSpec.durableApproval` / `Checkpoint.durableApproval` are typed `DurableApprovalOptIn | false` (the row types that carry the member, `CheckpointRow.durableApproval` and `CheckpointFields.durableApproval`, read `false` too; the opt-in's `scope` / `ttlMs` members now sit on the named root type `DurableApprovalOptIn`, unchanged — `TaskSpec.durableApproval.scope` reads after a `!== false` narrowing); `ParkLaneReason`, `SuspendLaneReason` and `CheckpointErrorReason` each gain `leg_not_resumable`; a new row never carries the synchronous-child block. (2) Who is hurt: a host flow built on those cards (server S-1471 / S-3463: the 409 → 200 decide road of a synchronous child's row has no new row to decide) — today such a card only continued in the same process with the right config, its result never reached the parent, and after a restart it stayed pending forever; a TypeScript reader of `.scope` on `durableApproval` (a narrowing `!== false` first); an exhaustive switch over the three closed sets; a host that builds the Workflow tool itself and passes `parentDurableApproval` (its run now fails at entry, loudly). (3) What to do: work that must wait for a later approval goes to a background agent (its row can be resumed and its completion is notified); a host keeps no decide entry for a synchronous child's card, expires the rows the listing now marks `leg_not_resumable`, renames `parentDurableApproval` to `parentDurableApprovalForChildren` where it builds the Workflow tool, and adds `resume.parent_constraint_missing` (still raised for a background / workflow row resumed without its parent chain) and `leg_not_resumable` to its supplementary sentence table (@cli @client-core: neither reads either word today).
|
|
1158
|
+
- **Hard-breaking — the Workflow mount's opt-in dep is renamed (open to veto @server @sdk).** (1) What broke: `RunWorkflowToolDeps.parentDurableApproval` is now `parentDurableApprovalForChildren` (the name of its context twin `ToolExecuteContext.durableApprovalForChildren`); the old name is the new clamp row's reserved per-seat name, and `createRunWorkflowTool` refuses it `config.deps_retired_key` with a sentence naming the new key. (2) Who is affected: a host that builds the Workflow tool's deps itself and passes the old key (the engine's own mount is renamed with it; server / cli / client-core / sdk `src` read 0 occurrences). (3) How to migrate: rename the key; the value (`{ scope, ttlMs? }`) is unchanged.
|
|
1159
|
+
- **Not breaking — `RunWorkflowOptions.defaultDurableApproval` (a wire-manifest reading).** It is typed `Exclude<TaskSpec["durableApproval"], false>` where it was `TaskSpec["durableApproval"]`: the declared text changed because `TaskSpec.durableApproval` gained `false`, and the exclusion keeps the option what it was — `DurableApprovalOptIn` (the same `{ scope, ttlMs? }`) or absent; every earlier argument compiles, and a workflow default never says `false`.
|
|
1160
|
+
- **Narrowing (open to veto @server @cli).** On a synchronous / fork child, a background child whose lane cannot park, an observer sidecar and a retained agent's wake cycle: a safety-vocabulary ask, a question and an unanswered-by-a-live-seat permission ask are refused instead of parked as a dead card; a plan-review request is refused to the asking call instead of parked; a usage window / an env lifetime stops the leg loudly (`usage.window_exhausted` / `env.lifetime_expired`) instead of parking it (the parent's next boundary parks a resumable row); agents of a workflow run with no park lane (no run store) park nothing. A root leg may declare `false` (opt-in to the narrowing).
|
|
1161
|
+
- **Narrowing — the forced gate on a store-less leg nobody resumes (open to veto @server @cli @client-core; the same direction as C-R1388 note 2).** Under `forceDurableGate`, three kinds of leg used to put their asks to the live approver and now refuse them without asking anyone: a fork child (it runs store-less by design, #3989), a delegated leg whose run turned its store off (`checkpointStore: "disabled"`), and a child in a deployment that wired no store. The leg is read before the store: such a leg is announced `config.durable_gate_unavailable` with `cause: "leg_not_resumable"` (it read `task_store_disabled` / `no_deployment_store`), its live approver is asked 0 times, the action does not run, and the refusal names the way out. A root leg without a store is unchanged (`task_store_disabled` / `no_deployment_store`, answered live). What to do: give the work a leg whose parks can be resumed (a background agent on a lane that can park, or the parent leg itself), or do not set the entitlement for that principal.
|
|
1162
|
+
- **Wording (open to veto @server @cli @client-core).** The three asks such a leg cannot put to anyone read one refusal (`notResumableRefusal`): what wanted a person, "and this leg cannot be resumed — <the leg's cause> — so it was refused without asking anyone (fail-closed)", what came of it ("the call did not run (do not report it as done)" / "nobody answered it"), "do not retry it on this leg, where nothing can wait for a person", then the way out and — for a call — "Meanwhile, tell the user what was blocked and why you needed it, and continue with the parts of the task that need no approval." (a question keeps "Proceed with your best judgment and state the assumption you made."): the four things the reference's refusal of a call nobody can approve tells the model, in this engine's words. The question no longer says "(no checkpoint facility, or the suspend failed)" there (the store was there; the cause is the leg); an inherited durable mandate's refusal now names the leg's cause and the way out (it named neither); the leg's own mandated ask keeps its substance. The deny exit names the way out for every ask nobody was asked about (one seat, where the mandated ask spelled its own). A question put to a live face that gave no answer (no mandate keeps it off the face) says "so it was refused once the live face gave no answer" — "without asking anyone" is said only where nobody was asked (a forced leg, or no face) — and a question's way out says "answer" ("have a person who is present answer it live …", "its questions wait for a person to answer later"). Where the host declared `durableApproval: false`, every refusal on the leg names the host's cause and its way out, also on a lane leg (`notResumableSaysOf`: the host's `false` rides to every child, so "ask from the parent leg, or a background agent" named roads that are closed there); the park arms still read the lane first. What the parent reads is the next entry (`refusedAsks`).
|
|
1163
|
+
- **Widening (open to veto @server; a background lane that parks for a person where the reference's background agents refuse rather than ask is registered as D-4087-a, which this widening extends).** Under a forced principal, a background child on a lane that can park now synthesizes its own opt-in: behind a synchronous parent (it was refused everything) and on a host whose root runner resolves no entitlement while the child's does (it got no forwarded opt-in and was refused) — both now park, resumable (#4096 F5's shape, now one reading).
|
|
1164
|
+
- **Three questions.** Who needs it: an unattended deployment whose synchronous children met approvals (a dead card, a parent told "suspended" for good); a host whose lane cannot resume an approval (server S-3941, the leader's static lane: `false`). Who is hurt: the hosts above (hard-breaking (2)). Compensation: the refusal and the notice name the way out; the background lane keeps the waiting card with a real resume and a completion notice; the old rows are listed honestly instead of failing at the decision.
|
|
1165
|
+
- **Rules.** Readings of "this leg's park can be resumed": 4 → 1 (the proxy; the safety-vocabulary arm, the review lane and the plan pair that read only "a store") — every reader reads the one fact by name. The park closure's own forced decline retired (the closure is not built where nobody resumes). The synchronous-child block's mint, its saga seat and the resume's restore retired (C-R227 (d)). Added: one value (`false`), one lane flag (`parkLaneHeld`), one reason word per closed set (one word, three sets), one row-judge arm; one way-out sentence pair shared by every refusal of such a leg (`notResumableWayOut` / `notResumableClause`) and one refusal sentence for the asks it cannot put to anyone (`notResumableRefusal`; three spellings → one); the gate carries the leg fact (`ToolGateInput.notResumable`), not a composed sentence. The forced gate's refusal fact reads the leg only (the store term retired: "can this leg be resumed?" is asked before "is there a store?").
|
|
1166
|
+
- **Pins.** `test/backlog4111-park-only-on-resumable-leg.test.ts` (N1–N12 with N1b–N1d / N2b / N3b–N3d / N4b–N4f, C1, R1, F1–F3 the three asks' one refusal, G1–G5 the leg before the store; on base `fdfdf94a4` 22 red, the 5 controls N1b / N1c / N2 / N3c / N11 green on both; F1–F3 / G1–G2 red on `c1650cbba`, the controls G3 / G4 / G5 green on both); re-pinned with the ruling that moved them (C-R1405 K1 (a)): `test/backlog3789-parked-row-ask-id.test.ts` R6 / S1 / U4, `test/backlog4054-sync-child-question-one-judge.test.ts` ④ (+ ④b) and its question cells (a leg nobody resumes reads its own refusal; the #4012 closing words), `test/backlog4079-forced-gate-resumable-leg.test.ts` ⑨ (the park closure's decline; the refusal fact reads no store) and its question cells, `test/backlog733-durable-park-opt-in-declared.test.ts` (the fact's sources). `test/backlog515-p3-crash-delegation-approval.e2e.test.ts` (#515 P3, real SIGKILLs) was built on the synchronous child's park and splits along the two lanes that still carry its scenario: K1 kills a background child on a lane that can park while its approval is pending — the row and its checkpoint survive, nothing replays them, and recovery takes the lane's own road (`claimParkedAgent` + the parked-resume drive): the gated action runs exactly once, only after the decision, the row settles `completed` with its park residue cleared and one completion notification; K2 kills a synchronous delegation while its child runs — the parent's wake compensates the dangling Agent call (`interrupted_outcome_unknown`) and the re-issued delegation runs to its end. The re-present through a re-issued delegation is gone with the synchronous park (the lane holds the row, so the decision is taken on it).
|
|
1167
|
+
- **New optional members — what a leg did not get done (C-R1405 note 2; open to veto @server @cli @client-core @test).** `TaskResult.refusedAsks?: readonly RefusedAsk[]` (`RefusedAsk` = `{ toolName, toolCallId, cause: AskUnresolvableCause, remedy? }`, both types exported at the root): one entry per ask the run refused because nobody could answer it, in refusal order, minted at the gate's one deny exit beside the `delegation.ask_unresolvable` notice (the same facts; `cause` is that notice's closed set `ASK_UNRESOLVABLE_CAUSES`, `remedy` the way out the refusal named). Present only when at least one ask was refused — never an empty array; the run's terminal is unchanged. The same list rides the background row (`BackgroundAgentRecord.refusedAsks?`, written by the terminal settle, cleared on a revival) and the completion frame (`TaskNotificationPayload.refusedAsks?`, rendered as `<refused-asks>`), and the synchronous and fork report cards carry it as a `refused_asks:` field line; the line and the tag are one text (`refusedAsksText`): how many, each tool and its cause word, "nobody could answer it, so it was refused and did not run", and the ways out. A decoded frame whose `refusedAsks` is not a non-empty list of well-formed entries is refused like any other ill-formed facet. Not carried yet (KL-4111-c, #4268): a workflow's agent steps — the workflow's child rows, its `workflow_complete` frame and its full record (`WorkflowRun.recordArtifact`) carry no `refusedAsks`.
|
|
1168
|
+
- **Hard-breaking — the row of a background child on a lane that cannot park (open to veto @server @cli @client-core @test).** (1) What changed: such a child that meets an approval or a question used to end its row `failed` with `error` `suspended.awaiting_approval` / `suspended.needs_review` (its checkpoint left pending, its card dead); its row now ends `completed`, with `refusedAsks` saying which asks were refused, why, and the way out — the child ran to its end, a refused ask being one tool result it went on from (the same reading as a synchronous child). The two codes are no longer produced on that lane, and no pending row is left. (2) Who is hurt: a fleet view or a host keyed on `failed` + `suspended.*` for such rows; a test that pinned the two codes on it (the 7.67.0 criterion G2.b's delegation report `{ error: "suspended.awaiting_approval", checkpointId }` has no producer on a synchronous, fork or un-parkable lane any more). (3) What to do: read `refusedAsks` (row, frame or `TaskResult`) for "done, but these asks were refused"; work that must wait for an approval goes to a background agent on a lane that can park.
|
|
1169
|
+
- **Consumers.** A host that starts a leg itself with `RunInternals.isDelegatedChild` / `observerLeg` now gets a lane leg that parks nothing unless the same internals say `parkLaneHeld: true` (the lane resumes that leg's parks); the engine's own lanes set it. @server: S-3941 (leader static lane) declares `durableApproval: false`; S-1471 / S-3463 (the synchronous-child decide road) retire; FIX7151-CAPS's forced child cells read the new widening. @cli: decide whether `-p` legs declare `false` (INTEGRATION-CORE 4.6m-bis); add the two words to the supplementary sentence table. @client-core: the same table. @sdk: the type widening.
|
|
1170
|
+
|
|
1171
|
+
### A host's manual compaction request is taken or refused on the tick of the call: `TaskStream.inputWindow()` is the stream's input judge, and a `compact()` made when no turn boundary of the run can come any more is refused at once (`steering.not_running`) instead of being parked and answered `mooted` when the stream settles (#4204; C-R1438; server S-4069; behavior change: one window; narrowing: none; widening: none; hard-breaking: one, below, open to veto @server @client-core @cli; closed sets: none; exports: `InputWindow` (type) added at the root, none removed; wire: `TaskStream.inputWindow` added) (@server @client-core @cli)
|
|
1172
|
+
|
|
1173
|
+
- **Rule.** One judge decides, synchronously, whether a manual compaction request made now is taken: `TaskStream.inputWindow()` — the function the stream already had for its own relay, with its words `"before"` / `"open"` / `"over"` (`InputWindow`, now exported at the root). `compact()` reads it before its first await and reads nothing else for that question: `"over"` (no turn boundary of this run will come: the loop has ended, or the run has settled) ⇒ refused, the promise already rejected when `compact()` returns; `"before"` / `"open"` ⇒ taken, the promise is its outcome. Taken is the admission, not a promise of service: a run that ends before another boundary comes answers a taken request `mooted` (for an instant after the loop's last prompt settles and before its end is recorded, the judge reads `"before"`, and a request taken there ends that way). A host route answers its request from the judge (taken ⇒ 202, refused ⇒ 409), before or after the call, with zero ticks.
|
|
1174
|
+
- **Before** (measured on the 7.68.0 integration tree, real Runner — `test/backlog4204-compact-bounded-receipt.test.ts`). A `compact()` made after the loop's end and before the result (the teardown) was parked and resolved `mooted` with a `compaction_outcome{mooted, manual, task_ending}` frame at the stream's settle, the same tick as `done`; a route that read the first tick saw no refusal and answered 202 for a request that was taken and never served: it ended `mooted`, with that one frame, when the stream settled. A run whose preparation threw read `"before"` from its run body's end until its result landed.
|
|
1175
|
+
- **Now.** In that window `inputWindow()` reads `"over"` and `compact()` is refused on the tick of the call: `steering.not_running`, "cannot steer: the task has already finished" (the sentence it already used once the result was out), no frame — a refused request is not an outcome. A run body that settled without ever publishing a handle reads `"over"` from that moment. `mooted` now has one meaning: the request was taken and never served (its caller withdrew it, a stop / park / plan review cut the loop first, or the run ended before a boundary came).
|
|
1176
|
+
- **Hard-breaking — one, three sentences (open to veto @server @client-core @cli).** *What changed:* in the `"over"` window `compact()` used to resolve `mooted` (with its `compaction_outcome` frame) when the stream settled, and now rejects `steering.not_running` when it is called, with no frame. *Who is affected:* a caller that calls `compact()` during the run's teardown and handles no rejection — one that awaits it without a `catch`, and one that sends it and reads nothing (`void stream.compact()`: the promise it gets is already rejected when the call returns, so with no handler on it that is an unhandled rejection, which ends a Node process by default) —, and a host or shell that counts `mooted` answers or `compaction_outcome{mooted, manual}` frames as the record of a teardown-time request; at the type level, a host that builds its own `TaskStream` object, because `TaskStream.inputWindow` is a required method of the type (a compile error until it is added). *How to migrate:* read `inputWindow()` before calling (on `"over"`, compact the idle session with `Runner.compactSession` once the stream has settled), or put a `catch` for `steering.not_running` on the promise `compact()` returns — on a call that is not awaited too — the code `compact()` and `steer()` already threw once a run had finished; a host-built `TaskStream` adds `inputWindow()` answering its own `InputWindow` word.
|
|
1177
|
+
- **Divergence from the reference (D-4204-a, ruled C-R1438, open to veto @server @client-core @cli).** The reference queues a `/compact` submitted while a turn runs and never refuses one for want of a live turn; this engine refuses a request made in the `"over"` window on the tick of the call, because no turn boundary of the run can come to serve it.
|
|
1178
|
+
- **Not changed.** The six `CompactOutcome` words and which boundary answers a taken request; the cancel seam; coalescing and per-call instructions; `"before"`: a request made while the run is starting is taken, and while the run has published no handle it is refused late, `steering.not_running` "the task is not running": when the run's preparation threw, once the failed preparation has settled (the same tick as the failed run's `done`; measured 86 ms and, on a host at load ≈ 40, 8.3 s after the failure), and when the handle is still unpublished 30 000 ms after the call (`READY_TIMEOUT_MS`, `src/core/runner/runtask.ts:644`, the readiness race every stream verb awaits; G5 reads it exactly) — a preparation that is only slow and goes on to start is refused at that bound all the same, and the run then completes (G5b; a lone `steer()` is refused at the same instant — its bound is counted from when its own turn in the steer queue begins, so a second queued steer meets it one bound later). A request taken while the run's last boundary is being served, or while its stop gate decides, still ends `mooted` at the stream's settle when no further boundary comes (measured, G7). `steer()`, `interrupt()`, `halt()`, `destroy()`.
|
|
1179
|
+
- **Rules.** Retired: one — the compaction seat's private "run body over" latch (`ManualCompactRef.closed`); its fact is the judge's (`TaskStreamLiveSeat.bodySettled`). The compact verb no longer keeps its own liveness reading (it read the result, then the latch); it reads the stream's judge. Added: none — the public method is the existing function.
|
|
1180
|
+
- **Pins.** `test/backlog4204-compact-bounded-receipt.test.ts` — G1–G8, one per window, each asserting that the judge read at the call and the first-tick state of `compact()` agree (`"over"` ⇔ rejected; `"before"` / `"open"` ⇔ pending): G1 live mid-turn (`open`, then `compacted`) · G2 settled stream (`over`, refused) · G3 starting, preparation then succeeds (`before`, then `compacted`) · G4 starting, preparation throws (`before`, nothing while it is held, refused when the failed run settles; `over` after) · G5 starting, preparation wedged (`before`; on a fake clock: pending at 29 999 ms, refused at 30 000 ms) · G6 a second request while a pass is in flight (`open`, `noop` at the next boundary) · G7 a second request while the last boundary's pass is in flight (taken, `mooted` + frame at the settle) · G8 teardown (`over`, refused on the tick, zero `compaction_outcome` frames); J0 the judge's words at each moment, a run body settled without a handle included · J1 / J2 `compact()` follows the judge when the live seat says otherwise, both ways · J3 taken while starting, the loop over when readiness arrives ⇒ `mooted`, said on the stream, never parked. Re-pinned: `test/manual-compact-cancel.test.ts` ③ (the teardown request is refused, no frame) and ④'s reacting consumer (refused, where it read `mooted`); `test/design393-s4-stream-lanes.test.ts` (the compact verb's judge and the stream's `inputWindow` are the steer lane's function).
|
|
1181
|
+
- **Consumers.** @server (S-4069 / S-4044): its compact route (`src/http/routes/runs.ts`, read on the server's main) answers 202 as soon as the run is in its live map and calls `compact()` fire-and-forget, so a `steering.not_running` rejection in the teardown window is only logged (`manual_compact_not_running`) after the 202 went out; its 409 (`compact.not_running`) is for a run not in the map. Reading `inputWindow()` before answering gives 202 or 409 on the tick of the call, and the `run_terminal` known limit for that window can retire once the route does. @client-core (CC-1516) / @cli: the re-send decision — on `"over"` (or this refusal) compact the session once the stream has settled, rather than retrying the stream. A host that builds its own `TaskStream` object adds the method.
|
|
1182
|
+
- **Wire.** Added: `TaskStream.inputWindow` (a required method — readers additive) and the root type `InputWindow` (existing words). No closed set moves.
|
|
1183
|
+
|
|
1184
|
+
### What stands where a text was withheld is one table — `REDACTION_MARKS` in `@sema-agent/primitives`, fifteen members (`«redacted»`, `«redacted:<kind>»`, `«redacted:<kind>:partial»`), one reader, and every writer of the engine takes a member; the square-bracket spellings are retired (#3970; #3959 S1, ruled C-R1318; #3356 and #3979 close with it; Security; behavior change: every mark the engine writes changes bytes; narrowing: four, below, each open to veto; widening: one, below, open to veto; hard-breaking: twelve, below (open to veto @server @cli @client-core @sdk); closed sets: `REDACTION_MARKS` (an object table — not seen by gate:wire, disclosed here); exports: root −1 / +5, primitives −1 / +8; wire: every engine mark)
|
|
1185
|
+
|
|
1186
|
+
- **Rule.** One table, `REDACTION_MARKS` (`src/primitives/redaction-marks.ts`, re-exported from the core root): `bare` = `«redacted»`; `key`, `gh-token`, `aws-key`, `slack-token`, `private-key` = `«redacted:<kind>»` and each `«redacted:<kind>:partial»`; `userinfo`, `url`, `path`, `oversized` = `«redacted:<kind>»`. Ten closed words, one closed shape (`«redacted`, at most two `:`-led words of 1–16 `[a-z-]`, `»`; written `^«redacted(?::[a-z-]{1,16})?(?::[a-z-]{1,16})?»$`) checked when the table loads. One reader — `isRedactionMark(s)`, `redactionMarkerSpans(s)`, `eachRedactionMarkerSpan(s)` (lazy) — recognizes the fifteen members exactly; nothing else (`«redacted:evil»`, `[redacted]`) is a mark. One kind rule: where the text around a mark says what was withheld (a credential name's value, an assignment, a header value, a scheme word's credential, the password behind a kept username, a URL query or fragment value, a sent credential a remote echoed) the mark is bare; otherwise it names the kind (a credential read by its shape — `key` or the vendor's word; a URL userinfo withheld whole; a whole URL or path on the human tier; a stretch that could not be read). The fragment grade writes the `:partial` member. Every value family row carries a required `mark` (never a `:partial` one); a caller's `RedactionPass.marker` is typed `RedactionMark` and checked at the call, a string `replace` must write exactly that member (plus `$n` references), and a replacer function is checked on each hit — what it writes in place of the hit is one member, or the call throws `TypeError`. One atomic law, read by the one interval renderer (`renderWithheld` exports it; the request-echo eraser uses it, so the two renderers of #3979 are one): a hit whose withheld part is exactly a mark the input holds is no hit; a hit reaching into or lying inside a mark takes the whole mark; removals that overlap or touch write one mark — the widest removal's (of two over one region, the later pass's). A display cut (`cutAt`), the bounded summary (its bare-ellipsis fallback included) and the inline / fenced untrusted-text clippers (`inlineUntrusted`, which bounds an approval card's command and digest summaries, and the fence's body cap) move back to a mark's start (#3356); a mark written where nothing was removed, inside a mark, takes that mark; the authority stripper (`stripUrlCredentials`) leaves a userinfo that is exactly a mark.
|
|
1187
|
+
- **Before.** Six tables minted marks and ten readers recognized them across the repositories; in the engine: `REDACTION_MARKS` held four square-bracket spellings (`[redacted]`, `[redacted-credentials]`, `[redacted-url]`, `[redacted-path]`), the URL display wrote `«redacted»`, the renderer took a caller's `SecretNotation` (`{ value, byKind?, byFamily?, overStrip }`), the reader matched the four bracket spellings only (so `cutAt("abc «redacted» def", 8)` read `abc «red`), the request-echo eraser kept a sent credential spelled inside a mark (`eraseSentCredentials("echo [redacted] end", ["edact"])` stayed as written), and a mark the input held was rewritten (`password=«redacted»` → `password=[redacted]`). Every credential was `[redacted]` whatever its kind; a URL password behind a kept username and a userinfo withheld whole were both `[redacted-credentials]`; an unreadable text was the notation's `overStrip`.
|
|
1188
|
+
- **Now (single layer, synthetic fills).** `use ghp_… now` → `use «redacted:gh-token» now` (AKIA → `aws-key`, `xoxb-` → `slack-token`, PEM → `private-key`, `sk-` / `AIza` / `hf_` / `glpat-` / JWT … → `key`) · `password=Hunter2…` → `password=«redacted»` · `Authorization: Bearer …` → `Authorization: «redacted»` · `https://alice:pw@h` → `https://alice:«redacted»@h` · `ftp://a@b@h` and a bare username of 20 characters or more → `«redacted:userinfo»@` · human tier: a URL → `«redacted:url»`, an absolute path → `«redacted:path»` · more than 64 format-character places, or percent layers that ran out → `«redacted:oversized»` · the fragment grade → `«redacted:<kind>:partial»` · `password=«redacted:key»` stays as written, `password=«redacted»TAIL42` → `password=«redacted»` · `cutAt("abc «redacted» def", 8)` → `abc ` · the engine's other writers: a credential-named member of a value walk or a terminal detail → `«redacted»`, the caller's-text shape every refusal sentence writes for a host's or declarer's text (`callerText`: a store's options bag, a retention knob, a host floor member, an MCP / peer seat member, a setting …) → `«redacted» (N characters)`, a URL display's query and fragment values → `«redacted»`, a withheld userinfo in a route refusal, `stripUrlCredentials` and the unparseable-URL display arm → `«redacted:userinfo»`, the sync peer's refusals → `«redacted»`, the request-echo eraser → `«redacted»` for a declared value or cookie value and the row's member for a value the vocabulary found (an over-large matcher → `«redacted:oversized»`).
|
|
1189
|
+
- **Old spelling → member, for a reader of logs, error lines and stored text.** One old spelling can become several members: the kind rule picks by what the text around the mark says. Rows marked *measured* ran the same synthetic inputs through `redactSecrets`, `redactHostLeaks`, `stripUrlCredentials`, `callerText` and `eraseSentCredentials` on the `v7.67.1` tree and on this release's tree; the others read the two trees' code. Every refusal sentence that echoes a host's or a declarer's own text goes through `callerText` (52 call sites in 17 judge modules) and changes with it. A reader that matched the bracket spellings reads the members through `isRedactionMark` / `redactionMarkerSpans`, never by a prefix or a pattern of its own.
|
|
1190
|
+
|
|
1191
|
+
| Old spelling (≤ 7.67.1) | Where the engine wrote it | Member now | Example, before → after |
|
|
1192
|
+
|---|---|---|---|
|
|
1193
|
+
| `[redacted]` | a credential the text around it names: an assignment, a header value, a scheme word's credential, a URL query value, a credential-named member of a value walk or a terminal detail, a sent credential a remote echoed | `«redacted»` | `password=[redacted]` → `password=«redacted»`; `Authorization: [redacted]` → `Authorization: «redacted»`; `?client_secret=[redacted]&page=2` → `?client_secret=«redacted»&page=2` (measured) |
|
|
1194
|
+
| `[redacted]` | a credential read by its shape with nothing around it naming it | `«redacted:gh-token»`, `«redacted:aws-key»`, `«redacted:slack-token»`, `«redacted:private-key»`, `«redacted:key»` | `use [redacted] now` → `use «redacted:gh-token» now` (a `ghp_` token), `use «redacted:aws-key» now` (an `AKIA` id), `use «redacted:slack-token» now` (`xoxb-`), `use «redacted:key» now` (an `sk-proj-` key); a PEM block → `«redacted:private-key»` (measured) |
|
|
1195
|
+
| `[redacted]` | a credential residue the fragment grade found | `«redacted:<kind>:partial»` (five members) | `cut [redacted] here` → `cut «redacted:gh-token:partial» here` (code: the old notation wrote `[redacted]` for every credential) |
|
|
1196
|
+
| `[redacted]` | a text no reading could afford (more than 64 format-character places; percent layers that ran out) | `«redacted:oversized»` | seventy `x` + U+200D pairs then `y` → `[redacted]` → `«redacted:oversized»` (measured) |
|
|
1197
|
+
| `[redacted] (N characters)` | `callerText`: a host's or declarer's own text in a refusal sentence | `«redacted» (N characters)` | `[redacted] (15 characters)` → `«redacted» (15 characters)` (measured) |
|
|
1198
|
+
| `[redacted-credentials]` | a URL password behind a kept username | `«redacted»` | `https://alice:[redacted-credentials]@host.example/repo` → `https://alice:«redacted»@host.example/repo` (measured) |
|
|
1199
|
+
| `[redacted-credentials]` | a URL userinfo withheld whole (a second `@`, a bare username of 20 characters or more; `stripUrlCredentials`) | `«redacted:userinfo»` | `https://[redacted-credentials]@host.example/repo` → `https://«redacted:userinfo»@host.example/repo` (measured) |
|
|
1200
|
+
| `[redacted-url]` | the human tier: a whole URL | `«redacted:url»` | `see [redacted-url] for it` → `see «redacted:url» for it` (measured) |
|
|
1201
|
+
| `[redacted-path]` | the human tier: an absolute path | `«redacted:path»` | `read [redacted-path] now` → `read «redacted:path» now` (measured) |
|
|
1202
|
+
| `«redacted»` (not a member then) | a URL display's query and fragment values | `«redacted»` (bytes unchanged; now a member the reader recognizes) | `?sig=«redacted»` → `?sig=«redacted»` (code) |
|
|
1203
|
+
| `«redacted»@` (not a member then) | a withheld URL user in a route display or refusal, the unparseable-URL display arm | `«redacted:userinfo»@` | `https://«redacted»@host.example/…` → `https://«redacted:userinfo»@host.example/…` (code) |
|
|
1204
|
+
|
|
1205
|
+
- **Hard-breaking — twelve, three sentences each (open to veto @server @cli @client-core @sdk).**
|
|
1206
|
+
- (HB-1, @server, API / type) *What changed:* the root `REDACTION_MARK` is removed; `REDACTION_MARKS` keeps its name but its keys and values change (`whole` / `credentials` are gone, `url` / `path` change value, eleven names are new); `RedactionFinding.marker` is typed `RedactionMark`; the primitives `renderSecrets` lost its notation parameter, `SecretNotation` is gone, `ValueFamilyRow.mark` is required and `RedactionPass.marker` is `RedactionMark`. *Who is affected:* the server's three importers move — `src/trace/redact.ts:43` imports the removed `REDACTION_MARK` and stops compiling (its tests `test/helpers/redaction-mark.ts` and `test/redact-engine-marked-scheme.test.ts` import it too, the second pinning it to `"[redacted]"`), while `src/trace/exit-seat.ts:39` and `src/orchestration/workflow-notify-journal.ts:43` import `REDACTION_MARKS` and read its `url` / `path` keys, whose values change (read from the server's origin/main on 11-13, not compiled here) — and its own closed-shape reader (`REDACTION_TOKEN_RE_SRC`) and the `ENGINE_MARKS` look-back (`src/trace/exit-seat.ts:160-183`) become a second reader. *What to do:* import the members and the one reader from the core root; read "a candidate that is exactly a member is no candidate" in place of "an engine mark is an already-withheld value", keep the scheme-word position rule but stop it looking at marks, delete the URL-display delegation probe, write members for the self-minted spellings (`oversized-diagnostic` / `-artifact` → `oversized`, a whole userinfo `«redacted»@` → `«redacted:userinfo»@`), and cut the recovery summary through the reader.
|
|
1207
|
+
- (HB-2, @server, persistence) *What changed:* the engine no longer writes the bracket spellings and the reader does not recognize them; nothing is migrated. *Who is affected:* rows a server store wrote before this release (sessions, journals, checkpoints) read back their bracket spellings as ordinary text, and the server's own `«redacted:oversized-diagnostic»` / `«redacted:oversized-artifact»` rows hold no member. *What to do:* do not reject, rewrite or recognize the old spellings; read the server's other `«…»` rows as members (KNOWN-LIMITS KL-3970-a).
|
|
1208
|
+
- (HB-3, @server, wire) *What changed:* the bytes of every mark the engine mints change (table above). *Who is affected:* the wire contract's reading sentence that recognizes engine marks by family, its engine-mark columns, and every server pin that holds an engine `[redacted]`. *What to do:* restate the contract as "recognized by member", point the columns at the table, re-pin by the reader, and correct the earlier CHANGELOG sentence that named the bracket spelling (pickup PICK768).
|
|
1209
|
+
- (HB-4, @cli, API / type) *What changed:* the same export changes as HB-1. *Who is affected:* the shell's own readers (`DISPLAY_REDACTION_MARKER_RE`, `src/displaySafeUrl.ts:330`; a `startsWith('«redacted')` at `:1556`) are second readers, its self-minted `:secret` / `:query` / `:fragment` words are no members, and `doctor` writes `[redacted]` (`modelRow.ts:804`). *What to do:* read through the core reader, write the bare member at the positions that say what was withheld (`:secret` / `:query` / `:fragment`) and `key` where nothing does, keep `userinfo`, and write the bare member in `doctor`.
|
|
1210
|
+
- (HB-5, @cli, persistence) *What changed:* nothing the shell stores carries an engine mark it reads back as one (inferred from the shell's stores; it keeps transcripts as text). *Who is affected:* a person paging an old transcript sees `[redacted]` as plain text, not styled as a mark. *What to do:* nothing — an old transcript is displayed as written.
|
|
1211
|
+
- (HB-6, @cli, wire) *What changed:* the human-readable output and `--output-format json` carry the new bytes as soon as the shell installs this release (its range `^7.64.0` takes it). *Who is affected:* every shell fixture that pins an engine `[redacted]`. *What to do:* re-pin them by the reader, not by spelling.
|
|
1212
|
+
- (HB-7, @client-core, API / type) *What changed:* the same export changes as HB-1. *Who is affected:* its four readers (`REDACTION_MARKER_PREFIX` `:59`, `MARKER_WITH_DOTS` `:237`, `MARKER_LEAD` `:316`, `MARKER_ANYWHERE` `:911`; `terminalToSdkResult.ts:350`), its self-minted `:secret` / `:query` / `:fragment` words, the `KIND_BY_TEXT` / `TEXT_BY_KIND` back-reading of a kind from text, and its reading of the retired `***` placeholder (a second reading that already exists). *What to do:* depend on `@sema-agent/primitives` (core is a devDependency today) and read through its reader, write members by the kind rule (bare at the three named positions, `key` elsewhere, `userinfo` kept), carry a span's kind with the span instead of reading it back from text, and retire the `***` reading.
|
|
1213
|
+
- (HB-8, @client-core, persistence) *What changed:* nothing persistent (inferred: the library stores no mark). *Who is affected:* no one. *What to do:* nothing.
|
|
1214
|
+
- (HB-9, @client-core, wire) *What changed:* the bytes it displays change with the engine's. *Who is affected:* its display fixtures that hold engine bracket spellings. *What to do:* re-pin them by the reader.
|
|
1215
|
+
- (HB-10, @sdk, API / type) *What changed:* the export changes of HB-1. *Who is affected:* no code reader (every hit in its source is documentation). *What to do:* nothing in code.
|
|
1216
|
+
- (HB-11, @sdk, persistence) *What changed:* nothing. *Who is affected:* no one. *What to do:* nothing.
|
|
1217
|
+
- (HB-12, @sdk, wire) *What changed:* the bytes its documentation quotes. *Who is affected:* `types.ts:1201` (`"[redacted]"`) and `:4091` (`[redacted] (N characters)`) now describe bytes the engine no longer writes. *What to do:* write `«redacted»` and `«redacted» (N characters)` there.
|
|
1218
|
+
- **@test.** Fixtures that pin an engine `[redacted]` (the c148c1 G7card reading `[redacted]«redacted:key»`) re-pin by the reader; from this release a card carries one grammar.
|
|
1219
|
+
- **@ss.** Its `[redacted-secret]` (`config-fns.ts:59-70`, `:132-142`, `:408`) is a sixth spelling: write the bare member, `key` or `oversized` by the kind rule and depend on `@sema-agent/primitives` for them.
|
|
1220
|
+
- **Narrowing (open to veto @server @cli @client-core).** The reader no longer recognizes the four bracket spellings: a text holding them is ordinary text to every reader — a display cut can land inside one, the pre-existing-marker count does not count it, and a strong credential name's value spelled `[redacted]` is withheld again (`password=[redacted]` → `password=«redacted»`, one finding).
|
|
1221
|
+
- **Narrowing (open to veto @server @client-core).** Two removals that touch write one mark even when they would write different members (the widest; of two over one region, the later pass's). #3958 kept touching removals of different marks apart; that reading is replaced, because two marks side by side behind a kept label read as one value to a second rendering, which then rewrites them. Kept: a later pass that overwrote each of two touching earlier marks exactly leaves two marks (the sequential reading).
|
|
1222
|
+
- **Narrowing (open to veto @server).** `«redacted:private-key:partial»` read again by the egress tiers becomes `«redacted»`: its words hold the strong name `private-key` and a separator, so a hit reaches into the mark and takes it whole; no byte of a credential is involved, and the other fourteen members are fixed points of both tiers (KNOWN-LIMITS KL-3970-b).
|
|
1223
|
+
- **Narrowing (open to veto @server).** A caller's pass whose replacer function writes anything but one member where it withholds text (a spelling of its own, two marks, nothing) now refuses the call with `TypeError`; before, the text it wrote was rendered and the finding named the pass's marker.
|
|
1224
|
+
- **Known (KNOWN-LIMITS KL-3970-c, KL-3970-d).** A shape-read credential behind a name writes the member of the pass that took it last: a strong name rewrites the vendor's mark bare (`password="sk_live_…"` → `password="«redacted»"`), a weak name leaves it (`api_key="sk_live_…"` → `api_key="«redacted:key»"`, `GITHUB_TOKEN=ghp_…` → `GITHUB_TOKEN=«redacted:gh-token»`). The region withheld is the same; which pass's member one region keeps is #3958's merge law, not changed here. And (KL-3970-d) the request-echo eraser run again over its own output, with a declared value that is a piece of a mark's words (`red`, `key`), takes that mark whole and writes it bare once; the third run changes nothing.
|
|
1225
|
+
- **Not done here (a sweep, offered as a follow-up).** About thirty other seats cut text that may hold marks with a raw or code-point slice (provider error excerpts, task and workflow previews, the activity-argument summary, the shared `sliceHeadSafe` / `sliceTailSafe`, tool-result previews); a cut there can still end inside a mark (no credential byte is involved).
|
|
1226
|
+
- **Widening (open to veto @server @cli @client-core).** A mark now names the kind of credential it withholds where the text does not (`gh-token`, `aws-key`, `slack-token`, `private-key`, `key`) and says when a userinfo was withheld whole: a reader of the output learns which vendor's credential stood there. No byte of the credential is shown.
|
|
1227
|
+
- **Not changed.** What is withheld from a text that holds no mark (the regions and the findings' spans; except the touching merge above — read off the re-pinned suites, whose region expectations held while their marks changed), the scan's grades and readings, `scanSecrets`, the egress tiers, the findings channel's members other than `marker`'s type, the `[not recorded: …]` notes (a separate glyph family), and the renderer's cost on many hits: the check that a hit cut into a mark reads only the marks the hit meets, so the #1583 cost input (16 000 hits behind one format character) renders in the time it took before.
|
|
1228
|
+
- **Divergence (D-3959-a, ruled C-R1318, open to veto).** The reference writes a square-bracket family of at least seven spellings that differ in case and separator and share no grammar a reader can parse; this engine writes the fifteen members of one table in one closed shape — the shape the server's contract already states on the wire.
|
|
1229
|
+
- **Rules (abstract review).** Engine side: spellings 5 → 15 members of one table, grammars 2 → 1, mint tables 3 (`REDACTION_MARKS`, `REDACTION_NOTATION`, `WITHHELD_VALUE_MARK`) → 1, readers 3 (`REDACTION_MARKER_RE`, `cutAt`'s four words, the echo eraser's "a hit inside a mark is none") → 1, renderers 2 (`markedOver`, `renderRemovals`) → 1; retired: `SecretNotation` and its validation, the `byKind` / `credentials` mapping, the "inside a mark is no hit" half rule; added: the `mark` column, the `RedactionMark` type, "`:partial` by grade". Two mechanisms carry the atomic law through the sequential renderer: the URL userinfo grammar reads a mark as one token (its `:` never splits a user from a password; the withholding cut and the `userinfoHolds` predicate read the userinfo's colon through that one reader), and a sequential run whose hit cut into a mark is re-rendered from its removals. Across the repositories (design census): 24 spellings / 2 grammars / 6 mint tables / 10 readers → 15 / 1 / 1 / 1. No special case, no transitional arm, no double read.
|
|
1230
|
+
- **Exports.** Root: −`REDACTION_MARK`; `REDACTION_MARKS` (same name, new members); +`isRedactionMark`, `eachRedactionMarkerSpan`, `redactionMarkerSpans`, `RedactionMark`, `RedactionMarkName`; `RedactionFinding.marker`: `string` → `RedactionMark`. `@sema-agent/primitives`: −`SecretNotation`; `renderSecrets(text, policy, options?)` (the notation parameter is gone); `ValueFamilyRow.mark` required; `RedactionPass.marker` / `RedactionFinding.marker`: `RedactionMark`; +`REDACTION_MARKS`, `isRedactionMark`, `eachRedactionMarkerSpan`, `redactionMarkerSpans`, `renderWithheld`, `WithheldSpan`, `RedactionMark`, `RedactionMarkName`.
|
|
1231
|
+
- **Pins.** `test/backlog3970-redaction-mark.test.ts` (51 cells; G1 the table · G2 the reader · G3 each row's member · G4 `oversized` · G5 the atomic law and fixed points · G6 cuts · G7 one mint · G8 the marker type and refusal · G9 the export surfaces · G10 the fragment grade · G11 the echo eraser); `test/backlog3958-overlap-redaction.test.ts` re-pinned to members (A7 by the touching merge). Every engine pin that held a bracket spelling is re-pinned to the member the kind rule writes, the caller's-text, provider-sentence and quoted-userinfo pins included (`backlog3186` / `3202` / `3826` / `1824-2130` / `3806` and 27 more).
|
|
1232
|
+
- **Consumers.** Matrix (repository · writes · reads · by import or literal · reading the old spellings):
|
|
1233
|
+
|
|
1234
|
+
| Repository | Writes | Reads | By | Old spellings |
|
|
1235
|
+
|---|---|---|---|---|
|
|
1236
|
+
| core | the table's four words + the URL display's `«redacted»` | `cutAt`, the echo eraser's needle ranges, the pre-existing-marker count and `boundedString` | the table | a cut could split an old mark; the count does not count one |
|
|
1237
|
+
| server | 13 + 2 self-minted; the engine's url / path spans (`exit-seat.ts:186-191`), the journal (`workflow-notify-journal.ts:145-150`) | the closed shape (three uses), two engine-origin special readings | engine marks by import, its own by literal | an old `"[redacted]"` credential field is rewritten to `"«redacted»"`; the rest stays; the scheme-word position rule is the exception (inferred) |
|
|
1238
|
+
| cli | 4 words + `doctor`'s `[redacted]` | the closed shape (`:330`), a prefix (`:1556`) | literal | an old spelling is text and is washed again (the safe side; inferred) |
|
|
1239
|
+
| client-core | 4 words | four readers + the kind back-reading (`:221-234`) | literal | the same (inferred) |
|
|
1240
|
+
| sdk | none | none (documentation `types.ts:1201` / `:4091`) | — | — |
|
|
1241
|
+
| ss | `[redacted-secret]` | none | literal | — |
|
|
1242
|
+
| test | fixtures | fixtures | literal | a fixture reds and is re-pinned |
|
|
1243
|
+
|
|
1244
|
+
### Two functions a store twin copied by hand are on the package root — `staleRunningFlipOf` (the form a stale running background-agent row flips to), `staleRunningFlip` (the sweep's whole judgement) and `compareTaskIds` (the task list's `list` order) (#4256; server request; additive, no behavior change)
|
|
1245
|
+
|
|
1246
|
+
- **Rule.** One writing of each fact: the flip's form is `staleRunningFlipOf`, which `staleRunningFlip` (the in-memory and file stores' reap arm 1) now calls instead of writing the members itself; the list order is `compareTaskIds`, which the in-memory store sorts with.
|
|
1247
|
+
- **Now.** `staleRunningFlipOf(live, now): BackgroundAgentRecord` — pure, a copy: `status: "failed"`, `stoppedBy: "system"`, `summary` / `error` = `STALE_RUNNING_REAP_ATTRIBUTION` where the row records none, `settledAt` = `updatedAt` = `now`, `completionId` minted fill-once; every other member, `rev` included, is the live row's. It judges nothing (the caller found the row running, readable and stale — a SQL twin's staleness is a column predicate). `staleRunningFlip(live, key, scope, now, staleMs, disclose?)` is the judgement (scope, running, reads, `now - updatedAt > staleMs`) and answers that form with `rev + 1`. `compareTaskIds(a, b)`: numeric ids ascending, any other id after them, lexicographically, ties told apart by text.
|
|
1248
|
+
- **Not changed.** What either bundled store's `reap` writes; the order of the in-memory store.
|
|
1249
|
+
- **Consumers.** @server: `flippedForm` in the SQL background-agent store is a hand copy of the form (its 7.67 miss: the `mintCompletionId` line) — it can call `staleRunningFlipOf(live, now)` and keep its own CAS / `rev` placement; the SQL task-list store's order can sort with `compareTaskIds`.
|
|
1250
|
+
- **Wire.** Root exports +3 (functions: `compareTaskIds`, `staleRunningFlip`, `staleRunningFlipOf`); none removed, no type or signature moved (`wire-changes` 3 added / 0 BREAKING).
|
|
1251
|
+
- **Pins.** `test/backlog4256-flip-form-and-task-id-order-root-exports.test.ts` (R root identity; F the form member by member, input untouched; J the judgement and its single writing; S both bundled stores' `reap`; O the order).
|
|
1252
|
+
|
|
1253
|
+
### The file task-list store lists in `compareTaskIds` order for an id that is not a canonical integer, and the task-list store contract kit now pins it (#4256; behavior change for such ids; contract narrowing, open to veto @server)
|
|
1254
|
+
|
|
1255
|
+
- **Before.** `createFileTaskListStore(...).list()` read the document's object key order: canonical integer ids ascending, every other id in write order. Written `30 b 4 a 12 007 7 99999999999 5000000000`, it listed `4 7 12 30 b a 007 99999999999 5000000000`; the in-memory store (and the `list` contract) gave `4 007 7 12 30 5000000000 99999999999 a b`.
|
|
1256
|
+
- **Now.** The file store sorts its listing with `compareTaskIds`: `4 007 7 12 30 5000000000 99999999999 a b` — the one order every backend lists in. Engine-minted ids are `String(counter)` from 1 (`allocateId`), so a list the engine alone wrote reads the same as before; the difference shows only for an id a host or an outside writer chose (leading zeros, not all digits, longer than an array index).
|
|
1257
|
+
- **Contract kit (narrowing, open to veto @server).** `taskListStoreContract` gains one cell, "list order: an id that is not a canonical integer is placed by `compareTaskIds`". A host's own store that lists such ids in its backend's order goes red there; @server's SQL task-list store (it reports one existing order difference from the engine store) runs against the cell at pickup.
|
|
1258
|
+
- **Pins.** The kit cell (run by the memory and file backends in `test/mailbox-tasklist-stores.test.ts`); `test/backlog4256-flip-form-and-task-id-order-root-exports.test.ts` O6 (file store, including a reopened store).
|
|
1259
|
+
|
|
1260
|
+
### A `memory.target_outside_store` refusal says which of its four cases it is — `detail.outsideReason`, one closed set — and the person's `memory.mount_failed` notice carries that word and says that case's step; a store re-anchored under an open session now says "open a new session", not "re-anchor"; `reanchorMemoryStoreRoot` takes the root a person confirmed, `expectedFrom`, and records nothing when the anchor no longer records it (#4210; behavior change: one member on every refusal of the code, one optional member on the notice, one refusal sentence reworded, one option added; narrowing: none; widening: none; hard-breaking: none; closed sets: `TargetOutsideStoreReason` new (`act_outside` / `root_moved` / `root_gone` / `root_reanchored`); exports: +2 (`TARGET_OUTSIDE_STORE_REASONS`, type `TargetOutsideStoreReason`; tier advanced); wire: `StoreFormRefusalDetail.outsideReason?`, `memory.mount_failed` `detail.outsideReason?`, `reanchorMemoryStoreRoot` option `expectedFrom?` — all additive) (@server)
|
|
1261
|
+
|
|
1262
|
+
- **Before.** One code, four cases with four remedies: an act whose directory is a link out of the store (replace the link), the store root resolving elsewhere than its anchor records (put it back, or re-anchor), nothing at the recorded root (put it back, re-anchor, or start over), and a session reaching a store re-anchored since it mounted (nothing to fix). A caller told them apart only by which optional members `detail` carried, and the person's `memory.mount_failed` sentence said one step for all four. The last case was minted by the moved-root seat, so its sentence told the reader to call `reanchorMemoryStoreRoot` — on a store that had just been re-anchored on purpose. And a re-anchor re-pointed the anchor wherever the directory resolved at the call: a refusal read before a second operator's re-anchor, handed back after it, re-pointed the store again.
|
|
1263
|
+
- **Now.** Every refusal of `memory.target_outside_store` carries `detail.outsideReason`, given by its own seat (`layout.ts`): `storeActRefusal` → `act_outside`, `storeRootMovedRefusal` → `root_moved`, `storeRootAbsentRefusal` → `root_gone`, and a new seat `storeRootReanchoredRefusal` → `root_reanchored`, which the engine's handle check (`assertRootAnchored(handle)`) and the file backend's session-root check (`assertRootAnchored(expected)`) now call. Its sentence states the mismatch, not the history behind it: "the store is anchored to another directory than the one this session opened it at — refused, nothing in the store was written; the store needs nothing: open a new session, which opens the store where it is anchored now". `controlDir` / `recorded` / `now` stay data on every case but `act_outside`, as before. `memory.mount_failed` reads the refusal once, through the family's one reader beside its one mint (`storeFormRefusalReadingOf`, `layout.ts`): `detail: { sessionId, scopes, cause, outsideReason? }`, `outsideReason` present exactly when `cause` is `memory.target_outside_store`; a value carrying that code without one of the four words reads `unclassified`. The person's sentence for each word is one table (`satisfies Record<TargetOutsideStoreReason, …>`); for `root_reanchored` it says "the memory this conversation opened has been moved to another place since … To restore it: start a new conversation, which opens it in its new place. Nothing else needs doing." — no operator step. `reanchorMemoryStoreRoot(dir, { controlDir, expectedFrom })`: `expectedFrom` is a refusal's `detail.recorded`, handed back as it came; under the anchor's lock and before the re-anchor writes it is compared character for character with the physical root the anchor records — another root, or an anchor absent or unreadable, refuses `config.memory_reanchor` and records nothing (the latter's sentence: to replace a damaged anchor, call without `expectedFrom`). Finding the store's control plane comes first either way, as at every open — a plane that travelled with a moved root is brought to its key — so the anchor compared is the one every opener reads. Without it the call is as before.
|
|
1264
|
+
- **Not changed.** The code, its membership of `STORE_FORM_REFUSAL_CODES` and `NON_GOVERNANCE_MEMORY_CODES`, the `act_outside` / `root_moved` / `root_gone` sentences as this section found them (#4145, below, rewrites the root-moved and root-gone ones to name the arguments by what they are), every other `detail` member, when each case is refused, the notice's minting, dedup and audience, and every other cause's sentence.
|
|
1265
|
+
- **Rules.** Readings of "which case": 1 (the word, from the seat that minted it) where a caller had to infer it from which members were present; readers of the family off a thrown value: 1, beside its mint (the notice's own reading of the code moved there); seats of the code 3 → 4, the fourth replacing a misuse of the moved-root seat; special cases: 0. The re-anchor gains one precondition, judged in the one place the anchor is read under its lock.
|
|
1266
|
+
- **Pins.** `test/backlog4210-store-refusal-reason.test.ts`: each case on a real File store — the word, the refusal's own step, the person's sentence for that word composed by the memory phase's reader (`act_outside` through a topic link; `root_moved` and `root_gone` at a new opener; `root_reanchored` at the engine's call site through the session's residue sweep and harvest, and at the file backend's call site through the session's commit root, directly and inside the harvest's lock wait); a real Runner leg whose store moved or went away hears `outsideReason` on `memory.mount_failed`; `expectedFrom` equal / a stale refusal after another re-anchor / anchor damaged and absent / not passed. Re-pinned: `backlog1103` (the refusal helper reads the word; ⑧ the detail's keys; ⑪a and ⑪c now read the re-anchored sentence), `backlog1094` and `backlog1368` (each detail carries its word), `backlog1409` (four seats, four words, no word on another code), `backlog2446` and `backlog4144-4146` (the notice per case; and a stale `expectedFrom` as the first call after an engine root moved: the travelled plane is found as every open finds it, the re-anchor records nothing, the root the next refusal names goes through).
|
|
1267
|
+
- **Consumers.** @server (the service's two asks): branch on `detail.outsideReason` — `root_moved` / `root_gone` lead to a re-anchor, now with `expectedFrom: detail.recorded` (the root the person saw), `root_reanchored` to a new session, `act_outside` to replacing the link; the HTTP body may carry the word without any path. `memory.mount_failed` `detail.outsideReason` is the same word. A refusal minted by an engine before 7.68.0 carries no word: read its absence as "an older engine", never as a case.
|
|
1268
|
+
|
|
1269
|
+
### A directory-pinned store under the config root keeps its control plane when the config root moves: an open finds the plane that travelled with the root and moves it onto the key the directory's path has now, and a moved store that holds memory is refused loudly instead of opening an empty plane; the key stays the one every build computes, so builds of different versions over one directory read one plane (#4146; user-visible fix — a declared "do not record me" was silently dropped; narrowing open to veto @server; closed sets: `STORE_FORM_REFUSAL_CODES` +1 (`memory.control_plane_split`), `NON_GOVERNANCE_MEMORY_CODES` +2; hard-breaking: one, the store-form refusal codes and the `memory.mount_failed` causes gain a member, open to veto @server; @test)
|
|
1270
|
+
|
|
1271
|
+
- **Rule.** A control plane belongs to the directory it serves, not to where the engine root sits, and every build that opens one directory opens one plane (C-R1423 note 2): one key formula, unchanged; a plane that travelled with a moved root is found and moved, never silently replaced by a new one (a bad value is loud).
|
|
1272
|
+
- **Before.** `deriveControlPlaneDir(configRoot, dir)` = `<configRoot>/<slug of dir's absolute path>/.engine` (`layout.ts`) — still the formula. A host that keeps its memory directory under its data root (a service's local file engine: `deriveControlPlaneDir(root, <root>/memory)`) and moves the data root got a key it had never seen: a fresh, empty plane with no anchor read as a first open, so the store opened over it. Capture opt-outs, the ledger and quarantine stayed behind in the plane that travelled under the old path's key, the session status answered `captureOptedOut: false` for a session that had declared it, and a resumed leg wrote its account row into the new plane. The "store root moved ⇒ refused" arm never fired, because its anchor lived in the plane nobody read.
|
|
1273
|
+
- **Now.** The key is unchanged. `effectiveControlRootOf` (the reading every opener shares; the explicit `controlDir` arm too, when that directory recomputes to the derivation, so a host that passes the derivation needs nothing) looks, at each open of a directory under the root and before the anchor head creates the plane or its lock, for the plane that travelled with the root: its key ends with the slug of the directory's path relative to the root; its root anchor reads, and the spelling it records mints exactly that key; nothing stands at the directory that anchor records, nor at the root that directory was under (a path that cannot be looked at stands; a plane at such a key whose anchor cannot be read refuses the open, `ControlPlaneCorruptError` — whether it holds the store's opt-outs is not known). One ⇒ it is moved onto the current key — in one rename while nothing but empty directories stands there (racing openers attempt the same rename, the loser reads the winner's); else record by record into the plane that an open by a build that does not look made there: the records written once (capture opt-outs, pollution markers, holds, quarantine, import receipts, clearance custody, the store identity) first, each kept there at once, even when the open is then refused; a record both planes hold with different contents refuses the open, `memory.control_plane_split`, naming the two key directories and the records, before anything else moves; else everything else is brought over (a record held there byte for byte is dropped from the travelled plane) and the travelled root anchor replaces the one there. The travelled anchor then judges the open: refused (`memory.target_outside_store`, the re-anchor step in its sentence) while the store holds memory, recorded again (#4145) while it holds none. More than one travelled ⇒ refused, `memory.control_plane_split`, naming them. None ⇒ a new plane, and its first open over a directory that already holds memory discloses `memory.control_plane_new_over_memory` once (copied, cloned, moved from outside the root, or a root moved while its old directory still stands — it says where the earlier records are and how to keep them) and opens. A move that cannot be made refuses the open (`ControlPlaneCorruptError`) instead of opening a new plane. Contract `memory.control_plane.key`.
|
|
1274
|
+
- **Builds of different versions over one directory.** Every build keys a directory the same way, so several hosts on one machine over the machine-wide data root, or replicas mid-upgrade over one volume, read and write one plane: an opt-out recorded by either reads as declared on both (pins). The one cell this build cannot reach: after an engine root move, an earlier build that opens the moved directory before this build has opened it once keys a new, empty plane and reads a recorded opt-out as absent — that build's own behaviour, unchanged here (measured on v7.67.0: `captureOptedOut` false, no incident). Once this build has opened it, the plane is where the earlier build looks: that build refuses (`memory.target_outside_store`) until the store is re-anchored, then reads the opt-out (measured on v7.67.0).
|
|
1275
|
+
- **Three questions.** Who needs it: a person who declared "do not record me" before an operator moved the data root, and the operator, who now gets a refusal that says the step instead of a silent empty store. Who is harmed: an operator whose moved store holds memory meets a refused open until one re-anchor call (a store that moved without memory records its root again, #4145); and one whose moved directory an earlier build opened first, where both planes came to hold one record with different contents, meets a refused open until one copy of each named record is kept. Compensation: each refusal says its step; the records written once are carried before any refusal, so a declared opt-out reads as declared on every build from that open on; nothing moves silently, nothing is dropped.
|
|
1276
|
+
- **Narrowing (open to veto @server).** A store under the engine root that moved with it and holds memory is refused until re-anchored (it opened over an empty plane before). A new store-form refusal, `memory.control_plane_split`: one more cause of `memory.mount_failed`.
|
|
1277
|
+
- **Hard-breaking (open to veto @server).** What broke: `STORE_FORM_REFUSAL_CODES` has six members (it had five) and `memory.mount_failed`'s `detail.cause` has seven values (it had six) — `memory.control_plane_split` joins both. Who is affected: a host that switches exhaustively over either set, or keeps a table keyed by it (@server, the mount-failed causes; a `satisfies Record<StoreFormRefusalCode, …>` fails to compile). How to migrate: add the arm — the operator's words are the refusal's own sentence (it names the two key directories and the records, and the step: keep one copy of each, remove the other, reopen); a person-facing table can follow core's (`mount-failed-notice.ts`: "keep one of the two copies of each record the service log names and remove the other").
|
|
1278
|
+
- **Rules.** Key forms 1 → 1 (unchanged; no adoption window, no retirement line). Arms +1: the plane that travelled is found by one four-part test and moved — one rename, or record by record by class (written once first; a conflict refuses; the rest; the anchor last). Refusal codes +1 (`memory.control_plane_split`). Incident +1 (`memory.control_plane_new_over_memory`, `NON_GOVERNANCE_MEMORY_CODES`: a silently new plane is a person's declaration silently lost). The control-object directory map moves into `layout.ts`, one map read by the object store and the merge.
|
|
1279
|
+
- **Not changed.** The key and every path. The plane of a memory directory outside the config root (moved ⇒ a new plane, said once), the personal plane (`<root>/personal/.engine`), the project-identity plane (`proj-<id>`), the repo partition and its earlier-key moves (`memory.partition_split` is still theirs alone). The SQL memory legs have no local plane keyed by the memory directory, and keep opt-outs keyed by session.
|
|
1280
|
+
- **Consumers.** @server: your local file engine passes `deriveControlPlaneDir(root, memoryDir)` as `controlDir`, the same path as before — nothing to change there; add `memory.control_plane_split` to your exhaustive handling of the mount-failed causes and of the store-form refusals (the hard-breaking item above); a refusal of your engine's open reaches the person through your own error path, since core opens no directory-pinned project store inside a run's memory phase; MIGRATION: no layout change and no adoption window; a data root moved with a store that holds memory now refuses its open until `reanchorMemoryStoreRoot` (the refusal says the step); builds of different versions over one directory read one plane, except an earlier build that opens a moved directory before this one has (it reads an empty plane, as it always did). @test: rc7 ④.
|
|
1281
|
+
- **Wire.** No key, word or frame. `STORE_FORM_REFUSAL_CODES` +1 and `NON_GOVERNANCE_MEMORY_CODES` +2 (`memory.control_plane_new_over_memory`, `memory.control_plane_split`).
|
|
1282
|
+
- **Pins.** `test/backlog4144-4146-store-set-leaks.test.ts` #4146: the key is the 7.67.0 formula and an opt-out written through either opening reads through the other; root moved, this build first ⇒ moved in one rename, refused, re-anchored as said ⇒ opt-out true, one plane; root moved, an earlier build first (the plane it makes) ⇒ the written-once records merged, a conflicting document refuses naming both while the opt-outs already read as declared there, no conflict ⇒ merged whole; two travelled ⇒ refused naming both; a copy (old root still standing) ⇒ not found, new plane said once; an orphan of another store with the same tail, its root still standing ⇒ not found, its root gone ⇒ found (the measured risk); the two crash windows; four processes opening at once ⇒ one plane; the derived arm.
|
|
1283
|
+
|
|
1284
|
+
### An open's residue no longer blocks a personal store after its engine root moved; a store that holds memory is still refused, with the re-anchor step said in words a caller handed only the sentence can follow; the person's mount-failed notice says how to restore it (#4145; user-visible fix; hard-breaking: none; widening open to veto @server)
|
|
1285
|
+
|
|
1286
|
+
- **Rule.** An anchor guards a store's records. An anchor that records another root guards nothing when the mount holds no record (`storeHoldsRecord`: only what an open or a read leaves) and nothing stands at the root it records any more (the store moved away whole) (C-R1423 note 1: three conditions — the anchor names another root, the mount holds no record, nothing stands at the recorded root, a path that cannot be examined counting as standing; the moved arm only). A root re-pointed while its old directory stands is refused as before, because the old directory may hold what the anchor guards (codex: files never adopted into a ledger) and because the memory directory must never become a redirect handle.
|
|
1287
|
+
- **Before.** A host's resolution of a personal key, a named read, a refused erasure, or a status read with a host carrier left the personal mount's anchor and empty directories. After the operator moved the engine root, the first personal session ran memoryless and a personal-layer erasure answered `memory.target_outside_store`. The refusal's sentence said `call reanchorMemoryStoreRoot(detail.path, { controlDir: detail.controlDir })`, which a caller handed only the sentence (an HTTP body) cannot follow; the root-gone and `memory.control_plane_strays` sentences did the same.
|
|
1288
|
+
- **Now.** `judgeAgainstAnchor` (`layout.ts`) answers "guards nothing" when the root resolves elsewhere, the mount holds no record and nothing stands at the recorded root. The open then records the root where it resolves now, under the anchor's lock, and discloses `memory.root_anchor_rerecorded` once (the backfill disclosure's form). A store holding memory is refused as before. Its sentence names the arguments by what they are — "call reanchorMemoryStoreRoot with the store's directory and { controlDir: its control plane directory } (for a deployment's personal store, `personal/memory` and `personal/.engine` under the memory engine root)" — and so do the root-gone and strays sentences. `detail` is unchanged. `memory.mount_failed` (audience `user`) now ends "To restore it: <action>. Whoever runs this deployment can do this (the service log names the exact step); the next run mounts it again." The action comes from one table keyed by the causes (`memory.target_outside_store`, `memory.scope_home_not_directory`, `memory.control_plane_links_unsupported`, `memory.control_plane_strays`, `memory.store_root_sensitive`, `memory.control_plane_split` — #4146 above —, `unclassified`); a cause added without its action is a type error. The cause clause for the strays no longer says "control plane" or "governance".
|
|
1289
|
+
- **Not changed.** The root-absent arm (it still reads `controlPlaneHoldsStore`). The codes and audiences; the `detail` keys as this section found them (#4210, above, adds `outsideReason` to the refusal and to the mount-failed notice). The mount-failed notice's minting and dedup. The mount-failed notice itself shipped in 7.67.0 (#2446); test's rc7 reading of "no person-facing sentence" was taken on 7.66.1.
|
|
1290
|
+
- **Widening (open to veto @server).** A store that moved away whole — an engine root moved by its operator, nothing left at the old path — over a mount that holds no record opens instead of refusing; the anchor is re-recorded and the incident says which root it had named. A root re-pointed while its old directory stands is still refused. The deployment's session status read with a host capture carrier (`MemoryDeployment.sessionMemoryStatus`) opens the personal store, so under the same three conditions it re-records too: a read entrance that writes the anchor and creates the store's directory (open to veto @server). The personal store then holds nothing and gets no row in `planes`; before, that read answered the plane `optOutSource: "fault"`. A personal store whose anchor names a root where something stands still answers `fault`.
|
|
1291
|
+
- **Rules.** Arms +1 (the moved arm splits on one existing predicate, `storeHoldsRecord`); incident +1 (`memory.root_anchor_rerecorded`, `NON_GOVERNANCE_MEMORY_CODES`); sentences pointing at `detail` 3 → 0.
|
|
1292
|
+
- **Consumers.** @server: `memory.mount_failed` has no entry in your notice whitelist, so it does not reach a tenant. Register code `memory.mount_failed`, audience `user`, cause closed set above. Whether a refusal's `detail` goes into the HTTP body is yours; the sentences no longer depend on it. @test: rc7 ③ (res145 / res151 / statres).
|
|
1293
|
+
- **Wire.** `NON_GOVERNANCE_MEMORY_CODES` +1. Three refusal sentences and one notice sentence change their words; the codes do not, and this section adds no `detail` member (`outsideReason` is #4210's, above).
|
|
1294
|
+
- **Pins.** `test/backlog4144-4146-store-set-leaks.test.ts` #4145: four residue sources × (moved ⇒ re-recorded once, session mounts); a store with memory ⇒ refused, memoryless, the person told; the refusal's step followed ⇒ the entry readable; the two other sentences; the seven causes rendered with no host path, function name or internal word. `test/backlog1103-memory-root-anchor.test.ts` re-pinned (the remedy said by what its arguments are; no `detail.` pointer). `test/backlog3427-deployment-write-home.test.ts`: D5's personal store that cannot be opened is one whose anchor names a root where something stands (it was an anchor naming a root that is gone, which this rule now re-records); D5b pins the re-record at the status read's entrance.
|
|
1295
|
+
|
|
1296
|
+
### A read of the personal store, and a session that kept nothing, no longer make a directory-pinned deployment a two-store one: the export's store identity is a named artifact, and a harvest's account of why it admitted nothing stays on its report (#4144; user-visible fix; hard-breaking: none; narrowing open to veto @server)
|
|
1297
|
+
|
|
1298
|
+
- **Rule.** Reads, and a session that kept nothing, leave nothing the store set counts as a record (`memory.deployment.store_set`).
|
|
1299
|
+
- **Before.** An export of the personal layer carrying no entry minted `personal/.engine/store-identity.json`, a file outside the named artifacts. A session that declared `capture: "off"` while writing the personal layer through a host carrier queued "harvest admitted nothing: … opt-out" into that scope's notice queue (`announcements.json`). Either way the set read two stores, and a request naming no layer was refused (server: 400 `request.memory_layer`), restarts included. A status read naming no layer with a host carrier opened the personal store; it left only named artifacts but was not stated anywhere.
|
|
1300
|
+
- **Now.** `store-identity.json` is one of `STORE_OPEN_ARTIFACTS.files`: only the two export read faces mint it, and its reader (`readStoreIdentity`, `layout.ts`) is the one the export and the census read, so a damaged identity counts as a record and the deployment's open refuses with the same fault. The harvest's own account of why it admitted nothing (an opt-out, a record-store outage, a caller's `admitNothing` reason, the pre-commit flip or outage) stays on `HarvestReport.warnings` and is queued nowhere. Every other warning of the same harvest (a boundary sweep's containment, say) is queued as before. The status read's open is stated in the contract sentence.
|
|
1301
|
+
- **Not changed.** A capture opt-out record kept on the mount (the file carrier) is a record, as ruled. The open law. The status read's host-carrier row.
|
|
1302
|
+
- **Narrowing (open to veto @server).** A damaged `store-identity.json` on a personal mount that held no other record left the set at two stores that opened; it now puts the store in the set and the open throws `ControlPlaneCorruptError`, as a damaged anchor or registry does.
|
|
1303
|
+
- **Rules.** Named artifacts +1, the identity: a read's artifact, read by the store's own reader, no new reading. One class of report-only warning, carried by the existing scoped-warning seat.
|
|
1304
|
+
- **Consumers.** @server: the 400 after an empty export or an opted-out personal session goes away; a closure assertion that copies `STORE_OPEN_ARTIFACTS` gains `store-identity.json`. Model face: a later session's memory notices no longer carry an opted-out session's "harvest admitted nothing" line (it is on that harvest's report). @test: rc7 ①②.
|
|
1305
|
+
- **Wire.** `STORE_OPEN_ARTIFACTS.files` +1. No key, word or frame.
|
|
1306
|
+
- **Pins.** `test/backlog4144-4146-store-set-leaks.test.ts` #4144: the empty export (set 1, the identity stays the store's), the damaged identity (three shapes: record, the open's fault), the opted-out personal session (set 1, no file), the read-only control, a real warning still queued, the status read's named artifacts. `test/backlog3345-store-set-artifact-closure.test.ts`: the read walk includes the export; the end state includes the identity.
|
|
1307
|
+
|
|
1308
|
+
### A delegated leg's taint frame says when its tree does not know the root's occupant — `permission.cross_boundary_taint` adds `root.occupantUnknown: true` in `occupant`'s place (#4260; behavior change: none to any decision, record or notice — one frame shape gains one optional member; narrowing: none; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: `TraceEvent` arm `permission.cross_boundary_taint`, `root.occupantUnknown` optional, additive) (@server @cli @client-core)
|
|
1309
|
+
|
|
1310
|
+
- **Before.** A delegated leg files its frames under its tree's root, `root: { sessionId, occupant }`. When the tree did not know the occupant its root held when it was born — the root's row has no token (a row written before occupants were minted) or the root leg's read of it failed, or the root facts a host's internals or a revived row handed carried none — the frame was `root: { sessionId }`: the shape of a top-level leg's frame, which files under the leg's own session. A consumer could file it only by the session id, so a frame of an earlier tree whose root session was deleted and created again under the same id (a retained background child of that tree still running) could be filed with the later occupant's run.
|
|
1311
|
+
- **Now.** That frame is `root: { sessionId, occupantUnknown: true }` — in `occupant`'s place, never beside it. A top-level leg's frame (`{ sessionId }`), a delegated leg's frame whose tree knows the occupant (`{ sessionId, occupant }`) and a delegated leg's frame whose host handed no root (`{ sessionId }`, its own session, KL-3666-a) are unchanged byte for byte. One mint: `prepareCrossBoundaryLane` (`src/core/runner/cross-boundary-lane.ts`) reads `root` off the leg's notice route, as before; a destination whose occupant is no token by the engine's one occupant reading (`isOccupantToken`, the predicate under `sameOccupant`) carries the member.
|
|
1312
|
+
- **Reading.** `occupantUnknown` is the tree's root address with no token. The engine judges that same address, for the leg's notices and the frames routed to the root, by `sameOccupant(undefined, current)`: the same generation only as a current row with no token — a row deleted and created again always has one, so it is another generation. It is the word `delegation.delivery_discarded` already carries as `detail.addressed: "unknown"` for an address with no token.
|
|
1313
|
+
- **Consumers.** @server: a frame with `occupantUnknown` is judged by `sameOccupant(undefined, <the current row's occupant>)` — filed only with a current row that has no token, not filed otherwise (never filing it is stricter and safe too; the choice is yours). A frame without the member reads as before: `occupant` ⇒ that occupant; neither ⇒ the emitting leg's own session. @cli @client-core: a `TraceEvent` mirror, if one is kept, takes the optional member; no reader of the frame needs a change.
|
|
1314
|
+
- **Not changed.** Every other member of the frame, byte for byte (U5, R5); when frames are emitted; the lane's decisions and session records (the persisted record carries no `root`); the notices. The tree still reads "the root's row has no token" and "the root leg's read of it failed" as one (the root leg hands its children only a token), so one member says both.
|
|
1315
|
+
- **Rules.** Added: none — the frame now spells for one shape what its doc already said ("absent = unknown"), because on this frame absence also meant "top-level leg". Retired: none.
|
|
1316
|
+
- **Pins.** `test/backlog4260-taint-root-occupant-unknown.test.ts` — U1–U6 on the lane (a top-level leg carries neither member whatever root facts were in reach; a known occupant carries it and no mark; an unknown one carries the mark on prepare and observed; a delegated leg with no root, or whose root is its own session, carries neither; every other member identical; an empty-string occupant reads as unknown) · R1–R7 on a real Runner over a file session repo whose root row is this build's (a token) or one written before occupants were minted (no `incarnation` in its header): a top-level leg on the row with no token carries no mark; a foreground, a background and a fork child under a row with a token carry its occupant, and under a row with no token carry the mark (the foreground child's observed frame too), as does a child under a row whose occupant cell names nobody (R3b); a grandchild carries the tree root's mark; the child's frames differ in `root` alone; a durable resume of a delegated leg whose host hands a root with no occupant carries the mark, and with one carries it. Re-pinned (old → new `root`): `backlog3666` L3 and `backlog3676` R0, R2–R6, R8, `{ sessionId }` → `{ sessionId, occupantUnknown: true }`; their top-level and known-occupant cells are unchanged.
|
|
1317
|
+
- **Wire.** `node scripts/wire-changes.mjs`: `TraceEvent` gains the key `permission.cross_boundary_taint.root.occupantUnknown`, optional; breaking entries: none.
|
|
1318
|
+
|
|
1319
|
+
### A background agent's stop-cycle number has one home — its row — and one mint, in the claim that opens a cycle: a completion woken on the durable lane after the retained lane had run is no longer swallowed, and the number never falls back (#4234 S0; defect fix; behavior change: frame `seq` values, a parked drive's spawn frame moves behind its claim, the unretained first cycle's notification gains `seq: 1`; narrowing: one, below, open to veto; widening: none; hard-breaking: six type-level items, below; closed sets: none; exports: none added, none removed; wire: notification / fleet-frame `seq` values, below) (@server @cli @client-core @sdk @test)
|
|
1320
|
+
|
|
1321
|
+
- **Rule.** One row, one generation; one claim, one step; a number exists only once the claim that writes it has won. The row is the durable record when a store is wired, the registry handle otherwise (a synchronously born child: its steer handle). Every wake mints through `nextCycleSeq(row) = (row.seq ?? 1) + 1`, in the claim's own write: the retained fast lane (`claimTerminalRowForRevive`), the tier-3 and mailbox-drain revivals (`send-message-tool.ts`), the parked resume's consume flip, the rowless steer-handle wake. A fresh row is written with `seq: 1`. Every reader takes the number its claim wrote: an attempt reads it once, when its claim won, and every frame it sends carries that one reading.
|
|
1322
|
+
- **Before.** The retained fast lane's claim did not write the row's `seq` (only a settle and a tier-3 claim did), and the retain entry kept a counter of its own. A child woken on the fast lane and then, past the retain window, revived by a SendMessage re-minted a number the fast lane had already spoken: the parent run's injection queue dedups on `task_id:status:seq` for the run's life, so the revived cycle's completion was dropped while the injector answered `queued` — the parent model never saw it (measured: one fast wake then two revivals ⇒ the first revival lost; eight fast wakes then three revivals ⇒ all three lost). The host frames' `seq` fell back (9 → 2), which a consumer that folds by `seq` reads as a stale frame. Same shape on a drain revival, on a second writer's revival of the same row, and on the session park lane (`PendingSessionNotifications.pend`, which dedups on the same key until a drain).
|
|
1323
|
+
- **Now.** One reading per row across both lanes and every process: `seq` rises strictly across a row's cycles on the model lane, the host frames and the durable row (pins ① ② ④ ④b ⑧, counter-example (b)). A parked resume's drive sends its cycle's spawn frame after its consume flip wins; a drive that ends before the flip (a missing transcript, a refused floor, a lost flip) sends no spawn frame and names the cycle that parked on its terminal (⑨b). The resumed leg's ticks carry the flip's number (⑨c).
|
|
1324
|
+
- **Hard-breaking (1).** *What changed:* `SubagentRetainEntry.cycleSeq` is removed. *Who is hurt:* code that reads a retained child's cycle from the retain ledger (`SendMessageToolOptions.retain`); none in server, cli, client-core or sdk (grep of each repo's main, below). *How to migrate:* read `TaskRegistry.backgroundAgentCycleSeq(taskId)`, or the `seq` of the cycle's frames.
|
|
1325
|
+
- **Hard-breaking (2).** *What changed:* `TaskRegistry.settleBackgroundAgent`'s outcome loses `seq`. *Who is hurt:* a caller that passed `seq` (a type error on the literal); none outside core. *How to migrate:* drop it — the cycle's number is already on the row, written by the claim that opened the cycle; a settle closes a cycle, it never names one.
|
|
1326
|
+
- **Hard-breaking (3).** *What changed:* `TaskRegistry.parkBackgroundAgent`'s park loses `seq`. *Who is hurt:* a caller that passed it — @server's probe `test/core-probe-registry-policy.test.ts:237-239` parks a live handle with `seq: 7` and expects poll to answer 7 (a type error on the literal, and poll now answers the row's own number, 1); no source caller in any neighbour. *How to migrate:* drop it — a park opens no cycle; the parked row keeps its claim's number and the consume flip writes the next; expect poll to answer that number.
|
|
1327
|
+
- **Hard-breaking (4).** *What changed:* `SubagentRetainLedger.register`'s `entry` type no longer omits a `cycleSeq` member (the member is gone from `SubagentRetainEntry`). *Who is hurt:* nobody that type-checked before — an entry literal without `cycleSeq` is accepted exactly as before; reported because the signature text moved. *How to migrate:* nothing.
|
|
1328
|
+
- **Hard-breaking (5).** *What changed:* `RunInternals.afterCheckpointResolve`'s return type widens from `Promise<void>` to `Promise<void | { cycleSeq: number }>` — the post-consume hook may state the cycle its claim opened, and the runner runs the resumed leg as that cycle. *Who is hurt:* nobody that implements the hook (a hook returning `Promise<void>` is still assignable and leaves the leg's internals as handed in) — @server implements it (`src/http/resume-legs.ts:2438`, `commitBeforeLaunch`, resolving `undefined`) and is unaffected; a caller that READS the hook's result now sees the wider type. *How to migrate:* nothing for an implementer; a reader handles the object arm. @server: when your hook should state the cycle its claim opened, return `{ cycleSeq }` — today it states none, so its resumed legs run with the internals you hand in.
|
|
1329
|
+
- **Hard-breaking (6).** *What changed:* `TaskRegistry.consumeParkedFlip` returns `Promise<number | undefined>` instead of `Promise<boolean>` — the flip reports the number it wrote (the cycle its claim opened), `undefined` when it lost; the drive uses that return and never re-reads the registry, whose handle a second drive may have replaced while the flip was in flight (⑨d). *Who is hurt:* a caller that compares the result to `true` / `false` — @server's probe `test/core-probe-parked-registry.test.ts` at :157, :176, :192 and :212 (`expect(await REG.consumeParkedFlip(...)).toBe(true)`) reads red at each; no source caller in any neighbour (main grep: server's three source hits are comments, cli / client-core / sdk zero). *How to migrate:* test the result against `undefined` (`!== undefined` for a win), or assert the number (`toBe(<the row's seq after the flip>)`).
|
|
1330
|
+
- **Narrowing (open to veto @server @cli @client-core).** A background-agent row whose stored `seq` is not a stop-cycle number (a safe integer from 1: `0`, `-3`, `2.5`, `"3"`, …) now reads as damaged, exactly as a stored `rev` that is no count does: a listing answers it by its handle and verdict (member `seq`), `get` and every CAS refuse it in `agent_record.row_malformed`, the write gate refuses such a candidate, deletion by handle goes on. Before, the one mint stepped the garbage, silently: `0` minted `1` (the first cycle's dedup key — the woken cycle's completion swallowed, on both lanes since this change), `-3` sent `-2`, `2.5` sent `3.5`, `"3"` sent the string `"31"` on the wire. What a sender sees now (measured, the same answers a bad `rev` gets): the retained lane answers `steering.still_running`, tier-3 answers `not_found`. No engine version wrote such a value; an absent `seq` stays legal and wakes as `2`. @server: your Pg / TiDB row store decodes `record_json` with its own object reader, not `readStoredBackgroundAgent` (a root export, already in 7.67.1), so its rows do not get this reading — adopt the core reader to have it.
|
|
1331
|
+
- **Host-lane frame changes (@server).** (a) A parked resume's drive that fails before its claim (transcript gone, floor refused, flip lost) used to send a spawn frame and a terminal frame both naming `n+1` while the row stayed `n`; it now sends no spawn frame and a terminal naming `n`. A successful drive's spawn frame arrives after the consume flip instead of before the preflight. (b) A cycle's terminal frame takes its `seq` from the completion frame only (the internal `generation` fallback is gone); every background lane's completion frame now carries it. (c) `TaskRegistry.backgroundAgentCycleSeq` answers a parked row's own cycle (`n`), no longer the next one.
|
|
1332
|
+
- **Model-lane change (additive).** The completion notification of a background agent's first cycle carries `seq: 1` with or without a retain entry, and so does the fork lane's — its dedup key becomes `…:<status>:1` (the host frame already said 1). A server inbox fold sees one more key member on those frames.
|
|
1333
|
+
- **Not changed.** The retained wake's cap and refusal sentence (`resume.cap`, 8 per child), the tier-3 receipt (`details.seq` still the claimed cycle), the stale-cycle guard on settle (`cycle` stamps), poll's shape (both arms now answer the same number), the drain revival's rollback (a failed drain restores the pre-claim row, its number included).
|
|
1334
|
+
- **Rules.** Mints 6 → 1 (`nextCycleSeq`). Retired, with no second reader left: the retain entry's counter and its ±1 / take-the-larger adoption; the handle's forward realign plus its extra +1; settle's and park's `seq` input and write-back; the frame-side three-arm derivation (`cycleSeqOfFrames`) and the settle-time snapshot; the parked `+1` pre-read arm; the terminal frame's `generation` fallback; three hand-written `(x.seq ?? 1) + 1`. Added: the fast-lane claim writes `seq` (the same expression tier-3 uses); the runner reads the hook's stated cycle (one sentence: the number exists only once the hook's claim has won, the runner never computes one).
|
|
1335
|
+
- **Known limits.** KL-4234-a: a row whose fast-lane cycles ran on an older version reuses numbers after the upgrade — as many times as the last number the old process sent exceeds the row's `seq` (measured: the old process sent 1–4, the row said 1, three tier-3 revivals spoke 2, 3, 4 — all three reused); no migration, no dual read. Who folds the reused numbers: a cli process that stays online across the server upgrade (its arrival key carries `seq`); a cli that restarted rebuilds its keys from the transcript without `seq` and does not fold them (read in the cli and client-core code). KL-4234-b: a consume flip whose row write commits but answers an error reads as lost — the model sees the cycle fail at `n` while the row says running at `n+1` until the stale-running sweep; the next claim steps from `n+1`, so no number repeats (the attributed read-back is #4254). Not fixed here: two drives of one parked ticket racing while the flip is in flight still swap the handle under the first drive (#4253); the frames' numbers are right since (6).
|
|
1336
|
+
- **Consumers.** Neighbour code grep on each main (server `fbcb3ff18`, cli `a78ccb3d`, client-core `0e380814`, sdk `3708338`) for `settleBackgroundAgent` / `parkBackgroundAgent` / `backgroundAgentCycleSeq` / `afterCheckpointResolve` / `SubagentRetainEntry`: cli, client-core, sdk zero; server — source: the hook implementation above (unaffected); tests: one park with `seq` (item 3), the rest call these verbs without the retired members. @server @cli @client-core: a row's `seq` now rises strictly across both continuation lanes and across processes — your fold-by-`seq` and stale-frame readers need no change, and the revived terminals they used to fold away will appear. @sdk: type-level items only. @test: the pins below.
|
|
1337
|
+
- **Pins.** `test/backlog4234-one-generation.test.ts` — ① ② ③ ⑥ ⑧ a real Runner parent run (fast ×1 then tier-3 ×2 ⇒ 1–4 and the model sees all four; fast ×8 then tier-3 ×3 ⇒ 1–12; retain off ⇒ the first notification's `seq` 1 = its host frame's; no store ⇒ 1–3; a full mailbox's drain revival ⇒ 3), ④ two registries on one store (A fast 2, B tier-3 3, A fast again 4), ④b a second writer's drain revival ⇒ 3, ⑤ a tier-3 cycle that parks and is approved ⇒ 4, ⑦ a synchronously born child woken twice ⇒ 2, 3, ⑨a / ⑨b / ⑨c the parked drive's two arms and its leg's ticks, ⑨d two drives of one ticket with the flip's commit delayed (the first drive's frames and ticks carry the number its flip wrote, 2); counter-examples (a) a session-scoped child across two parent runs ⇒ 1–4, (b) the session park lane keeps every cycle's frame; the collection leg's complement cells C1 (a retained wake and a tier-3 revival on one row at once, ×3 ⇒ 1, 2, 3, the loser's answer an error), C2 (the retained cap: the 9th wake refused with `resume.cap` and no number moved; tier-3 after the window ⇒ 10), C4 (a stale terminal handle: tier-3 mints from the row ⇒ 4), C6 (no `seq` ⇒ 2; four non-numbers read as damaged), C15 (a lost flip: no spawn frame, a failed terminal at 1); `backlog1023-row-shape-reader` gains two `seq` forms through every face of the file store. Re-pinned (old reading → new, each with its reason in the file): `subagent-cycle-seq` (1) / (3) / (4), `backlog258-cycle-seq-outline` (5) / (6), `backlog1482-child-completion-record` C8a / C8b, `backlog3355-parked-child-terminal` S16, `subagent-child-gate` (4), and the flip's return (6) in `backlog3334-parked-resume-discharge`, `p1-3-completion-id`, `agent-park-legacy-resume-toctou-400`.
|
|
1338
|
+
|
|
1339
|
+
### A `workspace: null` leg's file history is keyed against the session's tree, so rewinds converge its edits where they were made — no longer the same-named file in the other checkout; a fork has no file history across a session tree — its own or its source's (#4280; data-loss fix; behavior change: every leg's file-history root is the session's root, a delegated child at its caller's root keys with its caller, the fork carry is made only between two sessions with no tree of their own, a fork that works in a tree of its own records and restores no file history, a declared root whose place cannot be read is refused whatever its role; narrowings: five, below, open to veto; widening: none; hard-breaking: none; closed sets: `onError(phase:"config").classification` +2 words, `file-history-outside-checkout`, `file-history-fork-checkout`; exports: none added, none removed; wire: none) (@server @cli @client-core)
|
|
1340
|
+
|
|
1341
|
+
- **Before (7.67.0 – 7.67.3).** A leg that sent `TaskSpec.workspace: null` on a session with its own tree (#3742) works in the base checkout and recorded its edits keyed against its own root, the base checkout. The two rewinds read those records against two different roots. The run's rewind (`rewindFilesTo` on a later leg that works in the tree) joined them on the tree — measured on the real Runner: the null leg edits `base/b.txt` `b` → `b-null`, the person has `tree/b.txt` = `tree-person`; the rewind past the null leg answered `complete`, wrote `tree/b.txt` = `b` (the person's file overwritten) and left `base/b.txt` at `b-null`; a tree file the null leg edited was not rewound (its key was relative to the base root, `.sema/worktrees/<name>/c.txt`). `Runner.rewindFiles` joined every record on the deployment's root, the base checkout (#4276, below): the null leg's own base-file records came back where they were made, while the tree legs' records were written into the base checkout. A null leg's own `rewindFilesTo` joined on its own root, the base checkout: the tree legs' records written there. Reachable when a host sends `workspace: null` on a session that declared its own tree and the history is live.
|
|
1342
|
+
- **Now.** Every leg's history is keyed against the session's root (`sessionRootOf`, the reading the out-of-run rewind uses — `openSessionCheckout` with no declaration): a null leg's base-checkout edits are recorded under their absolute path, its tree edits relative to the tree, and both rewinds join every record of this version on that one root, each converging where it was edited (rows recorded before this version: KL-4280-a below; history a fork carried from its source: the third narrowing below and KL-4280-d / KL-4280-e; a fork that works in a tree of its own: Narrowing 4). A child the null leg delegates that works where the leg works (its root is the leg's) keys against the leg's history root and records into the session's history, as a tree leg's child always has: a rewind to before the delegation restores the child's first touch too (measured, F3 below; without this, the child kept its own history and the rewind restored the leg's later bytes instead). Legs that work in the session's tree, sessions with no checkout, and their children key exactly as before.
|
|
1343
|
+
- **Narrowing 1 (open to veto): a `null` leg whose session's checkout cannot be resolved.** When the leg records or restores history and the session's tree is gone, no longer this session's, or unreadable, the leg is refused under the checkout's code (`workspace.checkout_missing` / `workspace.checkout_unavailable`), its sentence `TaskSpec.workspace is null on a leg of session "<id>", whose file history is keyed to the session's own checkout, and that checkout cannot be resolved — …`. Through 7.67.3 such a leg ran in the base checkout and recorded base-relative keys. A leg that neither records nor restores — no store, capture off (`RunnerDeps.fileHistoryCapture: "off"`), or `TaskSpec.fileHistory: "disabled"`, with no rewind asked — never reads the session's checkout and runs as before; a released checkout refused every leg before and still does; a session that never had a checkout reads the base root. A leg that mounts no tool is not in that list: whether a leg records is read from the store and the execution env alone (`src/core/runner/prepare-file-history.ts`, the `live` reading), not from its tools, so a `workspace: null` leg with `excludeAllTools: true` on a session whose tree is gone is refused `workspace.checkout_missing` like any recording leg, and the sentence speaks of "this leg's edits" though it can make none (the release review leg measured it: it ran on the tree before #4280 and is refused here; KNOWN-LIMITS KL-4280-f, #4397, 7.69.0). The way out today is the same as for any such leg: `TaskSpec.fileHistory: "disabled"` on it.
|
|
1344
|
+
- **Narrowing 2 (open to veto): a `null` leg whose root stands outside the session's checkout records no file history.** One reading for a null leg: its root is the session's key root when the session's trees hold it (its place in its tree), and outside otherwise — whatever the session listed, its task root or only a directory (a session with no tree keys every leg at its own root). A leg whose root the trees do not hold (the repository's parent, an unrelated directory, a directory in another repository) runs as before, but its edits are recorded nowhere: there is no root on which every later rewind of the session would join them back to the same file. Through 7.67.3 they were recorded against that leg's own root (a rewind sent from a leg at that same root restored them; the other rewinds joined them elsewhere). The operator hears it once per such leg on `onError` (`phase: "config"`, classification `file-history-outside-checkout`, the new word): `file history is not live for this leg — TaskSpec.workspace is null on a leg of session "<id>" whose root <root> stands outside the session's checkout, so this leg records no file history: its edits are not in the session's history and no rewind restores them. The session's trees: …`. A rewind sent on such a leg (`rewindFilesTo` / `resumeAt` + `restoreFiles`) is refused under the placement's code (`config.session_workspace_changed`), its sentence `TaskSpec.workspace is null on a leg of session "<id>" whose root <root> stands outside the session's checkout, so this leg records no file history and cannot rewind the session's files; the leg was not started. Send the rewind on a leg that omits TaskSpec.workspace, or through Runner.rewindFiles. …`. Retired when history rows carry their own root (#4283; KL-4280-c). A root whose place git gives no answer for is refused when unreadable, not treated as outside the repository: when the root cannot be canonicalized (other than as an absence: a path that resolves to nothing reads as in no repository), git does not run (an exec failure or a timeout), git exits non-zero while a `.git` entry stands at the root or above it (a repository is there and was not read), or git's answer is not a checkout top and a common directory, a leg that records history is refused `workspace.checkout_unavailable` (`TaskSpec.workspace is null on a leg of session "<id>": where its file history is keyed cannot be read, so the leg was not started. A leg that records no file history (TaskSpec.fileHistory: "disabled", or capture off) does not read it. The checkout's answer: Where this leg's task root <root> stands in git cannot be read (…)…`), nothing run; only a non-zero exit with no `.git` entry at or above the root reads as outside a repository (git's exit code is 128 for both, so the filesystem decides, never stderr text). Any failure on the way to reading the root — returned or thrown, whatever its code (an env adapter throwing `EIO` / `EACCES` from its path, file or exec call) — is refused the same way, not treated as outside the repository; so is a place in a tree that a link leads out of (under its own code, `config.session_workspace_invalid`). The tree legs' judgments (a root that holds an isolated checkout, two checkouts of one repository) are not a null leg's: such a root is outside like any other root the trees do not hold. A leg that records nothing (capture off) still runs.
|
|
1345
|
+
- **Code change on a refusal path (@server): a task root whose place cannot be read is refused `workspace.checkout_unavailable`, no longer `config.session_workspace_changed`.** A leg that works in the session's tree, and `Runner.rewindFiles`, whose task root git gives no answer for (the four shapes above; a step that throws surfaced as the thrown error, now refused the same way) were already refused at prepare time with nothing run, under `config.session_workspace_changed` (`This leg's the task root <root> is not in a repository this session's trees were minted from (…)`); they are now refused `workspace.checkout_unavailable` (`Where this leg's task root <root> stands in git cannot be read (…), so it is not placed in or outside the session's trees; the leg was not started.`, behind the verb's own sentence on `Runner.rewindFiles`): the code says the place was not read, not that the root moved. Both are refusals before any work; a host that answers by code sees the second where it saw the first, for this one cause only. A root that is read and stands outside the session's repositories keeps `config.session_workspace_changed`. The first leg's allocation reads its listed roots and its sources' roots through the same reader: one whose place cannot be read was refused `workspace.isolation_unavailable` (a listed root, "has no place in the session's tree") or `workspace.source_mismatch` (a source's root, "is not in a git checkout") and is now refused `workspace.checkout_unavailable` with the same sentence — still before any tree is made (pin `backlog3462` T13); a root git reads as in no repository keeps those codes.
|
|
1346
|
+
- **Isolation fix (narrowing, open to veto @server @cli @test): an additional directory whose place cannot be read is refused, not left standing as declared.** Every non-blank declared root — the task root, an `additionalDirectories` entry, an `additionalReadDirectories` entry — has its place read before the leg runs, on the first leg's allocation and on every later leg of a session with a tree; one whose place git gives no answer for (the shapes above: not run, a non-zero exit below a `.git`, an answer of no shape, a step that throws) refuses the leg `workspace.checkout_unavailable` (`Where this leg's additionalDirectories entry <dir> stands in git cannot be read (…), so it is not placed in or outside the session's trees; the leg was not started.`), nothing run, no tree made. What it closes: an additional directory in another (linked) checkout of the isolated repository, outside the base checkout, whose place read failed stood as declared and became a write root outside the session's tree — the second-checkout refusal above (#4009) reads that root only when its place is read. 7.67.1 let such a directory stand as a write root whether or not git answered (measured: the leg ran and wrote in the other checkout, on a later leg and on the first leg's allocation); in this version the read case is refused by #4009 and the unread case here. A directory that does not exist, or that git reads as in no repository, still stands as declared (a path that resolves to nothing is in no repository; the hands phase skips one that does not exist). A read-only entry is read the same way though it never becomes a write root, and three shapes of one that hold no repository are refused (the release review leg measured each on a later leg of a session with a tree: run on the tree before #4280, refused here; a plain directory and a missing path run on both): a path that names a file (its sentence says `git could not be run at <path>`, not that the path is a file), a directory holding a `.git` file that points nowhere, and a directory whose parent holds an empty `.git` directory (KNOWN-LIMITS KL-4280-g, #4398). A `null` leg declares no directory, so this does not reach it. One code moves with it: a writable root inside the isolated checkout's `.git` (git exits non-zero below a `.git` entry) was refused `config.session_workspace_invalid` and is now refused `workspace.checkout_unavailable`, before any tree (pin `backlog2938` I9).
|
|
1347
|
+
- **Narrowing 3 (open to veto @server @cli): a fork's copied history is carried only between two sessions that both key at the root their legs are handed — no tree of their own — and only when the fork's copy can vouch for that: on a store that wakes bounded windows, a fork woken from a window whose compactions were all written before 7.68.0 is not carried either, also when neither session ever had a tree (#3468 W14e).** A session forked by a host (`sessions.fork` + `runTask`, `warmResume`) copies its source's history on its first run or rewind (the fork carry); the records carry no root and were joined on the fork's root whatever root the source had keyed them against. Measured on the real Runner (fork of a session with no tree whose first leg declares its own tree; fork of a session with its own tree that works without one; the person's own bytes in both checkouts; a rewind to a point before the fork): the run's rewind wrote the source's bytes into the fork's tree (`tree-person` → `a`), `Runner.rewindFiles` and both rewinds of the second fork wrote them into the base checkout (`person-base` → `a`), each answering `complete`. Now the carry is made only when the fork's copy of its source's log vouches that the source held no tree of its own (`recordedRootOf`: a whole log with no checkout record of the source, or a record of it that isolates no tree — code sources only — reads as the root the source's legs were handed) and the fork keys at that same root. A copy that shows a record of the source that holds a tree carries nothing — the source worked in a tree of its own, or came to, and nothing on a history row says which of its rows were keyed there; nor does a copy with no readable record, or one woken as a bounded window in which no compaction restates the source's checkout (`SessionStorage.getFloorId`: a record below the floor is not ruled out; a compaction this version writes restates it for a source that is no fork, #3468, and such a copy vouches). The exception, a known limit: a copy cut from such a window as a new row has no floor to say what it cannot see and is carried (KNOWN-LIMITS KL-3468-a; #3468 W19w / W19p). Then the operator is told once on `onError` (`phase: "rewind"`, the store's `adopt_failed`: `file-history fork carry from session "<source>" was not made — <why>: …`, `<why>` = `the source's history is keyed against <root> and this session's against <root>`, or `this session's copy of the source's log shows the source working in a tree of its own, or cannot vouch for where its history is keyed`), and a rewind to a point before the fork is refused `rewind_snapshot.unresolvable`, whose sentence now says that a fork holds its source's boundaries only when neither session works in a tree of its own — the file changes before the fork are not in its history. Two sessions on the same deployment root with no tree of their own are carried as before (K3). **A fork placed in its source's tree (a host rooted in the source's own tree) is no longer carried** (K4: carried and restored on 7.67.1 and on the earlier cuts of this version; now refused, the tree and the base file left): the source's rows from before its tree were keyed against the base checkout, and a fork of a late-declaring source rooted in its tree wrote them over the person's tree file (K4L, measured: `tree-person-b` → `b`).
|
|
1348
|
+
- **Narrowing 4 (open to veto @server @cli): a fork that works in a tree of its own records no file history and restores none, in this version.** A session's key root can change only at its checkout's allocation (base checkout → its own tree), and only a fork's log can hold turns from before that allocation: a session that is no fork is refused a declaration once it holds a message or a compaction (#3468), while a fork may declare its own tree after it ran in the base checkout, or on a leg that branches (`resumeAt`) back above turns it ran there. History rows carry no root, so such a fork's history held rows keyed against the base checkout and rows keyed against its tree, and no reader can tell them apart. Measured on the real Runner (H2 / H2b / H2c: the fork records turns on the base checkout, then declares its tree on a leg that branches back — at a turn, `before` a turn, or after copying a source whose own abandoned branch is in the log; the person has their own bytes in the tree; a rewind to the abandoned turn): `rewindFilesTo` and `resumeAt` + `restoreFiles` answered `complete` and wrote the base checkout's bytes over the person's tree files on 7.67.1 and on every earlier cut of this version (`tree-person-a` → `a-S`), `Runner.rewindFiles` too on the earlier cuts (on 7.67.1 it restored the base checkout, where those turns ran), and its preview listed the files; a fork of such a fork cut before its tree carried the tree's rows onto the base checkout (H3: the person's `base/c.txt` overwritten). Now one judgment — a fork that isolated a tree of its own (`forkHasOwnTree`: the row's `forkedFrom`, and the session's own checkout record, read as every leg's open reads it, holds a tree — a cell its allocation isolated, released or not; a record that declares code sources only holds none and keeps the one root its legs are handed; `none` with a bounded window's floor is none) — closes that intersection: (1) such a session records no file history (no first-touch backup, no turn boundary; the leg's wiring fact `fileHistory: "disabled"`), said to the operator once on `onError` (`phase: "config"`, classification `file-history-fork-checkout`, the new word; once per session through the prepare ledger, which is the conversation's: a `resumeAt` back above the leg that said it says it again on that branch): `file history is not live for this session — session "<id>" is a fork that works in a tree of its own (TaskSpec.workspace): in this version such a session records no file history (no first-touch backup before an edit, no turn boundary) and a rewind of its files is refused; its edits in the tree are not recoverable by a rewind, and its conversation is still branched by resumeAt.`; (2) every rewind of its files is refused `rewind_snapshot.unresolvable` with the no-boundary sentence, nothing written — `rewindFilesTo`, `resumeAt` + `restoreFiles`, and `Runner.rewindFiles` preview and restore (after the checkout's own refusals: a released, gone or unreadable checkout keeps its code, C1). Its conversation is still branched by a `resumeAt` that asks for no files; a `resumeAt` with `restoreFiles: true` is refused whole — the refusal is thrown while the leg is prepared (`src/core/runner/prepare-file-history.ts`, before any restore) and the branch the request landed is written back (`src/core/runner/branch-writeback.ts`), so its conversation does not branch either (read from code). The operator line's last clause reads for the first form. A fork whose own record cannot be read when the judgment reads it is not judged to hold none: the leg and both rewinds are refused `workspace.checkout_unavailable`, nothing restored, recorded or written (`test/backlog3468-checkout-record-across-floor.test.ts` W18: a failed read injected at the judgment, the leg's open and the session root having read the record). What a person loses, each measured: a fork that declares its tree on its first leg cannot rewind its edits there (L2: refused, the tree keeps the new value; `complete` on the earlier cuts); nor can a late-declaring fork (L3 both targets, W16m after a compaction); its `workspace: null` leg's edit in the base checkout is not recorded and not rewound (C4); and **`Runner.rewindFiles` to a turn such a fork ran before its tree, which restored the base checkout on 7.67.1, is refused** (H2 / B3, verb). Sessions that are no fork, with or without a tree (L1, L1b, W14), and forks with no tree of their own (K3, W17: a fork of a compacted session, its own checkout read as none + window) record and restore as before. A child the fork delegates is handed no history lineage (the lineage is handed only by a run whose history is live), so it records into its own history as a child of a session with no history would. Not taken: history rows that carry their own key root (the final form, a separate ticket); a user-facing notice code (the refusal's sentence says why; none is added in this version).
|
|
1349
|
+
- **Known residual (KL-4280-a): rows a 7.67.0 – 7.67.3 null leg recorded.** They carry no root, and nothing recorded tells them apart (no engine version on a store row or a session entry, no record of the leg's `workspace`; a store row's `backupTime` is a clock reading, `workspace_state` is written only when the cwd changed): every rewind now joins them on the tree. Reach: every rewind of a session created on those versions (7.67.2 / 7.67.3 change none of this path: their `src` diff from 7.67.1 is five redaction files), with its own tree, that ran a null leg — whatever the target (before that leg, at it, or after it). Such a rewind (measured: O2, and the review leg's readings) writes the old bytes into the tree's same-named file of each base-checkout file the null leg edited (`tree/b.txt` = `tree-person` → `b`) and leaves `base/b.txt`; deletes a same-named file the person later created in the tree; leaves unrestored a tree file the null leg edited; and a rewind to the first turns can create a nested copy in the tree (`.sema/worktrees/<name>/c.txt` under the tree). `rewindFilesTo` already joined these rows on the tree in 7.67.x; `Runner.rewindFiles` restored them in the base checkout through 7.67.3. A preview lists only keys (`b.txt`), so it cannot show which checkout's file a row will write: for such a session, check the listed names against what the null leg edited before rewinding, or do not rewind it.
|
|
1350
|
+
- **Known residual (KL-4280-d): history a late-declaring fork recorded in its tree before this version.** A fork that ran in the base checkout, then declared its own tree and recorded there on 7.67.x or an earlier cut of this version, holds rows keyed against two roots in one history; a fork of it cut before that declaration reads "no checkout" for its source (its copy predates the record) and carries those rows onto the base checkout, and a rewind to the source's point answers `complete` and writes the tree's bytes over the base checkout's file (H3 on 7.67.1: the person's `base/c.txt`). From this version on such a fork records nothing in its tree (Narrowing 4), so no new history of that shape is written: only a history recorded before the upgrade is reached. Present on 7.67.1.
|
|
1351
|
+
- **Known residual (KL-4280-e): a host's deletion cascade carries a deleted source's history into its forks without the carry's judgment.** `RETENTION_CASCADE` step 10 has the host's coordinator call the store's `adoptScope` directly into each fork whose history is still empty, and step 11 deletes the source's; the rows land in the fork's history before its first leg, through no judgment of the fork carry, and a fork that works in the base checkout holds no checkout of its own, so Narrowing 4 does not reach it either. Measured with the two store calls the table states, made directly on the file store (not the host coordinator's reading): a source that worked in its own tree is deleted, its fork has recorded nothing and works in the base checkout, the person has their own `base/a.txt`; the fork's rewind to a point before the fork answers `complete` and writes the tree's original bytes over the person's base file, the tree left (pin `backlog4280-fork-checkout-off` CC). Present on 7.67.1 (the review leg's readings, same on every tree since). Not fixed here: the cascade would adopt only into forks the carry would carry, a change to the cascade table and its contract (and the host's coordinator, #4302).
|
|
1352
|
+
- **Not changed.** Where a leg works, what it may write, its frame (`workspace_changed`), the store, the request and answer shapes; the carry's address law (the fork row's lineage cell) and its empty-target law.
|
|
1353
|
+
- **Rules.** One root for a session's history across all its legs and both rewinds, read by one function (`sessionRootOf`); a child at its caller's root keys with its caller (the lineage's root), any other child as before; no file history across the intersection of a fork and a session tree — a fork that isolated a tree of its own records and restores none, and a fork's copied rows are carried only when its copy shows no tree of its source (one judgment function, `forkHasOwnTree`, read by the run's driver and the out-of-run rewind; the carry's reader is one expression; both read one predicate of a checkout record — whether it holds a tree — the one a leg that works in the session's trees reads); a null leg's root is the key root when the session's trees hold it, outside otherwise; a declared root whose place cannot be read refuses the leg; a leg reads the session's checkout for its history only when it records or restores. Retired: the walk from a fork's allocation up its ancestors (a point before the allocation), the carry's placement of the source's root through its trees, the rewind judge's read of the log's checkout entries. Added: two `onError` classification words.
|
|
1354
|
+
- **Pins.** `test/backlog4280-null-leg-history-root.test.ts` — N1–N4 × (`Runner.rewindFiles`, `rewindFilesTo`) · N5 a null leg's own `rewindFilesTo` · W1 the engine's own persisted cwd never moves the history root · F3 × (`Runner.rewindFiles`, `rewindFilesTo`): a null leg's sync child, its background child, and the two controls (a tree leg's child, a child in a session with no checkout) · C1 no checkout ever · C2 two repositories · C3 released · D1 narrowing 1, and the same shape with no store runs · O1 / O2 KL-4280-a's preview and real rewind. `test/backlog4280-history-root-edges.test.ts` — K1 / K2 × both rewinds: the two fork shapes above (nothing carried, refused, both checkouts byte for byte, one `adopt_failed` line naming both roots) · K6 a fork never run, rewound by the verb first (the verb's own carry) · K3 control (carried and restored) · K4 a fork placed in its source's tree (not carried, refused — Narrowing 3) · K5 the window reads as unknown · B3 / B3 control / F2 × both rewinds (Narrowing 4: refused, refused, restored without touching the person's file) · Z1 capture off: the null leg runs with the tree gone; asking a rewind it is refused · X1 × (the repository's parent, an unrelated directory): narrowing 2 end to end. Read in three trees (the 7.68.0 stage before #4276, the first #4280 cut, this one): N1 verb and N1–N4 run, N5, W1, D1 red before #4276 / #4280 (N2–N4 verb green there: the verb joined the base root, right for a null leg's base file); F3 null-sync / null-background green on the stage, red on the first cut (the rewind restored `B`, the leg's own first touch), green here; F3 tree-sync verb red on the stage only (#4276); the F3 and K3 / K4 controls green in all three; K1 / K2 / K6 red in both earlier trees (the fork's tree or the base checkout written); Z1 ran on the stage (its rewind joined on its own root), was refused on the first cut, runs here; X1 ran and recorded on the stage, was refused on the first cut, runs unrecorded here. `test/backlog4280-fork-checkout-off.test.ts` — A: H2 / H2b / H2c × (`rewindFilesTo`, `resumeAt` + `restoreFiles`, `Runner.rewindFiles` preview and restore) · C2 two isolated repositories · H3 · K4L · NR (the scope's keys and boundaries unchanged by two legs in the tree, one operator line, the fact `"disabled"`) · the refusal's sentence; B (the narrowings): L2 / L3 / C4 / C3 × both rewinds; C (unchanged): L1 / L1b / C7 × both rewinds, C1; residual: CC (KL-4280-e, pinned as measured). `test/backlog4280-fork-checkout-window.test.ts` — W17 / W16m / W14 × (the log window, the path window). The A cells answered `complete` (and wrote the wrong checkout) on the tree before this rule, the B cells answered `complete` there; mutations of each seat — the restore refusal, the capture switch, the carry's reader, the judgment without `forkedFrom`, a windowed `none` read as holding a checkout — each turn their cells red (car receipt).
|
|
1355
|
+
- **Consumers.** @server: `workspace: null` on a session-checkout session is safe to wire with this version; a session whose tree is gone refuses a null leg that records history (the two codes; a leg that mounts no tool records too, KL-4280-f); give `file-history-outside-checkout` its own log arm (a warning: edits no rewind restores) — it is not one of the two deployment-shape words; a host fork whose source worked in a tree of its own, or that works in a tree of its own, no longer has the source's pre-fork boundaries (a fork rooted in its source's tree included), and a rewind to one is refused `rewind_snapshot.unresolvable`; a host fork that declares a tree of its own (`TaskSpec.workspace` with `isolation`; a declaration of `sources` alone isolates nothing and is not one — #4395 below) records no file history and every rewind of its files is refused in this version — give `file-history-fork-checkout` a log arm of its own (a warning, like `file-history-outside-checkout`: that session's edits are in no rewind), and do not offer a file rewind on such a session; if your session store wakes bounded windows, a fork woken as a window that holds no restatement of its source's checkout (all its compactions written before this version) is not carried (a compaction this version writes restates the checkout of a source that is no fork, so a fork woken from its window can vouch for it — #3468, `test/backlog3468-checkout-record-across-floor.test.ts` W14e; a fork copied from such a window into a new row, with no floor, is carried — KL-3468-a, W19w; your SQL fork copies the whole stored log and does not reach it; a source that is itself a fork states no `none`, so its forks are not carried, W20c; KL-4280-e names the deletion cascade's opposite corner). @cli @client-core: "work outside the session tree" (a null leg) followed by "Rewind" restores the files that leg edited in the base checkout and leaves the tree's same-named files alone; a null leg on a directory outside the repository records nothing to rewind, and a rewind on it answers the sentence above; "Rewind" on a fork to a point before the fork, when either session works in a tree, answers "no boundary" instead of writing the source's files; "Rewind" of files on a fork that works in its own tree answers "no boundary" for every point — a conversation-only rewind (`resumeAt` without `restoreFiles`) still branches, one that asks for the files and the conversation together is refused whole and the conversation stays where it was.
|
|
1356
|
+
|
|
1357
|
+
### A fork whose workspace declaration names code sources only keeps its file history — the fork rule above reaches only a fork that isolated a tree of its own (#4395; correction to #4280 Narrowings 3 and 4 before release; behavior change against the earlier cuts of this version only — on a sources-only fork's own history every released version reads as now; the carry from a source is wider on 7.67.3, which also carries across a tree and across a bounded window's floor that this version refuses, #4280 Narrowing 3 and #3468; narrowing: none against the earlier cuts; widening: none against a released version; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none) (@server @cli @client-core)
|
|
1358
|
+
|
|
1359
|
+
- **Before (earlier cuts of this version, never released).** The fork judgment read "the fork has a checkout record of its own", and the carry's reader "the copy shows a checkout record of the source". A declaration may name `sources` and no `isolation` (`config.session_workspace_invalid` refuses only one that names neither); it mints a record whose cells carry no tree, every root stands where it was declared, and the session keys its history at the root its legs are handed — one root, as a session with no record. Such a fork was still judged: measured on the real Runner at the stage head (`247e56247`), a fork that declares sources only on its first leg and edits twice ran with `fileHistory: "disabled"`, heard `file-history-fork-checkout` once, and a rewind to its second turn was refused `rewind_snapshot.unresolvable` by `rewindFilesTo` and by `Runner.rewindFiles` (preview and restore), the edit left; its rewind to its source's turn was refused too; a fork of a source that declared sources only was not carried (`adopt_failed`: "this session's copy of the source's log shows the source working in a tree of its own, or cannot vouch …") and its rewind to the source's turn was refused. On the tree before #4280 (`b7c019c0a`) and on 7.67.3 the same inputs recorded and restored (`complete`, the files back).
|
|
1360
|
+
- **Now.** Both read one predicate of a checkout record, whether it holds a tree — a cell its allocation isolated (`dir`; a cell's four tree facts stand together or not at all), released or not — the reading a leg that works in the session's trees already made. A fork whose record declares sources only records and restores its history in both rewinds and every way of asking; a fork's copy of a source whose record declares sources only reads as a copy with no record reads — keyed at the root the fork's legs are handed (no session records the root its legs were handed: KL-4280-b) — and the carry is made when the fork keys there (a fork of such a source that isolates its own tree keys in that tree and is not carried).
|
|
1361
|
+
- **Not changed.** A fork whose record holds a tree — `isolation`, alone or with `sources`, released or not — records no history and restores none, and no carry crosses a tree (a released tree's copy included); a record that cannot be read refuses `workspace.checkout_unavailable`; a bounded window's floor that does not rule a record out carries nothing. A session's record never gains a tree later: a leg that declares `isolation` after a sources-only record is refused `config.session_workspace_changed`, on a leg that branches back above the declaration too (the reader scans the whole log), so a sources-only session's history never holds rows of two roots; a fork of a sources-only fork that declares its own tree is covered by the rule from its first leg. Sessions that are no fork, with or without sources, as before.
|
|
1362
|
+
- **Known residual (KL-4280-e) — back to its disclosed reach.** This correction gives a sources-only fork its file history and its rewinds back, and with them puts it back inside the disclosed cascade residual, the same shape as on every released version: the host's deletion cascade (`RETENTION_CASCADE` step 10) places a deleted source's history in a fork through no judgment of the carry, and a fork with no tree of its own — no declaration, or sources only — restores those rows, so a source that worked in a tree has the tree's original bytes written over the person's base file (measured with the two store calls made directly and the real Runner's rewind, pin CC). The earlier cuts' coarse rule refused every rewind of a sources-only fork and so kept it out of that reach for a while; the residual is one row, its condition now written whole. It reaches one hop further too: when the cascade has filled fork S and S is forked again into F, the hop S → F goes through the carry's reader and passes it (S's copy shows no tree of S's own), so F's rewind to a point before it writes the deleted source's tree bytes into the base checkout (the release review leg measured it with a sources-only S and with a record-less S, the same on 7.67.3). The close is the cascade's own (#4302) or rows that carry their root (#4328).
|
|
1363
|
+
- **Over-stop (KL-4280-h, P3).** A fork that isolates only an additional directory's repository and leaves its task root where it was still reads as a fork with a tree of its own and records no history, though its key root does not move (the release review leg: a session that is no fork with the same declaration records and restores its task-root edits). The judgment's grain is the record ("holds a tree"), not "the task root's repository is isolated"; the form is new in this version (7.67.3 refuses that declaration), so no released behaviour is lost.
|
|
1364
|
+
- **Rules.** The coarse rule covers only the case its reason names — a fork whose key root can have changed at its allocation; readings of "this record holds a tree" 3 → 1 (the leg's trees, the fork judgment and the carry's reader each had their own: a cell with `dir`, a record, no record). Renamed to say what they judge (internal, none exported): `forkHoldsCheckout` → `forkHasOwnTree`, the file-history phase input `forkWithCheckout` → `forkWithOwnTree`, `refuseRestoreOnForkWithCheckout` → `refuseRestoreOnForkWithOwnTree`. Added: none.
|
|
1365
|
+
- **Pins.** `test/backlog4395-fork-source-only-history.test.ts` (real Runner, real git, real file-history store) — A1 × (`rewindFilesTo`, `resumeAt` + `restoreFiles`, `Runner.rewindFiles`) a sources-only fork restores its own turn, every leg `on`, nothing said · A2 × 2 the same fork restores its source's turn (a source with no record) · A3 × 2 a fork of a sources-only source is carried · A4 both sources-only · A5 × 2 a sources-only fork of a source with a tree: its own turns restore, the source's turn refused, one `adopt_failed` · R1 a released sources-only source is carried · B1 × 3 / B2 × 2 a fork that isolates its task root (alone, at a declared source): `disabled`, said once, refused, the tree kept · R2 a released tree's copy carries nothing · D1 isolation after sources-only refused twice, no tree minted, history on · D2 × 2 a tree-declaring fork of a sources-only fork is covered from its first leg · N1 × 2 no fork · CC the cascade residual above, pinned as measured · U1 an unreadable own record refuses both entries `workspace.checkout_unavailable` · U2 the judgment and the carry's reader over the three record shapes. Mutations (car receipt): the predicate's two seats each read back as "has a record" ⇒ the A cells of that seat red; each read as "holds no tree" ⇒ the B / D2 / A5 / R2 cells of that seat red. `backlog4280-*`, `backlog4276-*` and `backlog3468-*` read as before (`backlog3468`'s injected read fault names the judgment's new name).
|
|
1366
|
+
- **Consumers.** @server: a host fork that declares `workspace.sources` without `isolation` keeps its file history and both rewinds in this version, as on 7.67.x; `file-history-fork-checkout` is said only for a fork that isolates a tree. @cli @client-core: "Rewind" of files on such a fork restores, as on 7.67.x.
|
|
1367
|
+
|
|
1368
|
+
### `Runner.rewindFiles` rewinds a session that works in its own tree in that tree — it no longer writes the base checkout and leaves the tree as it was; a checkout that cannot be resolved refuses the rewind (#4276; data-loss fix; behavior change: the preview and the restore read the session's root; narrowing: one, below; widening: none; hard-breaking: none; closed sets: none added (the session checkout's codes are now also thrown by `Runner.rewindFiles`); exports: none added, none removed; wire: none) (@server @client-core @cli)
|
|
1369
|
+
|
|
1370
|
+
- **Before (7.63.0 – 7.67.3).** The out-of-run rewind resolved its root from the deployment alone (`rootPath`, else the env's cwd) and never read the session's checkout, while a session's runs record their file history against the session's own tree (`TaskSpec.workspace` `isolation: "worktree"`), with keys relative to that root. On a session with its own tree, the preview diffed the base checkout and the restore wrote the recorded bytes into the base checkout — over whatever the person had there — answered `complete`, and left the session's tree unrewound. A released checkout or a deleted tree was not noticed either: the base checkout was written. Reachable when a session declares its own tree and a host calls `Runner.rewindFiles` on it; the run's own rewind (`rewindFilesTo`) on a leg that works in the tree was not affected (it reads the leg's root after the checkout is opened) — on a `workspace: null` leg it was, see #4280 above.
|
|
1371
|
+
- **Now.** The rewind reads the session's root the way a leg that omits `TaskSpec.workspace` does — the same `openSessionCheckout`, with no declaration, which never mints a tree and never writes the record: a session with its own tree is previewed and rewound in that tree (every repository's tree of a multi-root session), a session with no checkout at the deployment's root as before. The base checkout is never written in the tree's place.
|
|
1372
|
+
- **Narrowing.** A rewind whose session checkout cannot be resolved is refused, preview and restore alike — nothing previewed, restored or written (the refusal reads the session log, the session row, the tree's birth mark and runs git; no history-store verb is called) — under the code a leg would be refused with: `config.session_workspace_changed` (the checkout was released; also a deployment root outside the tree's repository — or `config.session_workspace_invalid` when that root holds the isolated checkout), `workspace.checkout_missing` (the tree is gone, or is no longer this session's), `workspace.checkout_unavailable` (the record or the tree cannot be read). Through 7.67.3 these answered `complete` after writing the base checkout. The message starts `Runner.rewindFiles: where session "<id>"'s files stand cannot be resolved …` and ends with the checkout's own sentence.
|
|
1373
|
+
- **Not changed.** A session without a workspace declaration (byte for byte); the request, the answer shapes, the admission order before the root as this release leaves it (the request door, the deployment door, busy, the session, the env, then `to` and the row's address — those two are this release's own, #3994 and #3996 above; 7.67.3 read neither); `rewindFilesTo`; the store contract.
|
|
1374
|
+
- **Known residual (narrowed by #3468 in the same release).** A store that wakes a bounded window (`SessionStorage.getFloorId`) reads the record from the newest compaction's restatement; only a window whose newest compaction did not restate the record, with the record below the floor, still reads "no checkout" for the rewind exactly as it does for a leg that omits the declaration, so such a rewind still reads the base root — a log whose compactions predate the restatement (KL-3468-a), a compaction whose read of the record failed or whose `details` is not an object (KL-3468-c), and a compaction made on a window that shows the release but not the allocation (KL-3468-d); the bundled full-load backends are not affected.
|
|
1375
|
+
- **Rules.** One reading of a session's root for both rewinds (the run's and the verb's) where there were two; no fallback arm; no code added.
|
|
1376
|
+
- **Pins.** `test/backlog4276-rewind-session-root.test.ts` — R1 the person's edit in the base checkout stays byte for byte, the tree is rewound, no session entry and no tree are written · R2 the preview reads the tree · R3 control: no declaration, base checkout as before · R4 two repositories, both trees rewound, neither base written · R5 released ⇒ `config.session_workspace_changed`, both trees untouched · R5b tree gone ⇒ `workspace.checkout_missing`, the base untouched · R6 control: `rewindFilesTo` on the same session. R1 / R2 / R4 / R5 / R5b are red on the base tree; R3 / R6 green by design.
|
|
1377
|
+
- **Consumers.** @server: the session-level checkout (S-1950) must not be wired to `Runner.rewindFiles` on a core older than this fix — before it, a rewind of such a session overwrites the base checkout; the four refusal codes above (`config.session_workspace_changed`, `config.session_workspace_invalid`, `workspace.checkout_missing`, `workspace.checkout_unavailable`) reach the rewind route and are configuration-class (not retryable as such; the released / missing forms have no rewind). @client-core @cli: a "Rewind files" action on a session with its own tree previews and restores that tree; the four codes are new answers of that action (the message says nothing was rewound).
|
|
1378
|
+
|
|
1379
|
+
### A session's checkout record and its workspace state survive a compaction floor: a store that wakes a session from its bounded window cut in log order reads what the whole log reads once the window's newest compaction restated them (KL-3468-a / -c / -d are the windows that do not) (#3468; behavior change; one refusal added to `Session.appendCompaction`, open to veto @server @cli; session-log data: two optional compaction keys)
|
|
1380
|
+
|
|
1381
|
+
- **Before.** A store that loads only `[floor .. leaf]` of a session (`SessionStorage.getFloorId`; the floor = the newest compaction's kept entry — the SQL backends' wake) loads no entry below the floor. A session's checkout record (`session_checkout`) is written once, on its first leg, so after a compaction whose kept tail started past it every wake read the session as having no checkout: a leg that sent nothing about the workspace ran in the base checkout and its edits landed there, not in the session's tree; a leg that sent the same declaration was refused `config.session_workspace_changed` ("already ran without it"); `releaseSessionCheckout` answered no trees and recorded no release; the orphan doors read the session as `record_unreadable`. The workspace state (`workspace_state`: the tracked working directory and an `EnterWorktree` tree) cut the same way started the next leg at the task root.
|
|
1382
|
+
- **Now.** Every compaction restates both facts on itself, written by `Session.appendCompaction` (so the run's passes, `Runner.compactSession` and a host's own call all carry them): `details.sessionCheckout` = what the record reader answered for the session — `{ sessionId, kind: "record", record }`, `{ sessionId, kind: "none" }` or `{ sessionId, kind: "unreadable" }`; `details.workspaceState` = the snapshot `getWorkspaceState` reads. The record reader takes a restatement of its own session as one more copy of the same record (every copy must hold the same allocation, a release anywhere releases it, a restated `unreadable` stays unreadable) — on every log: a bounded window, a log copied from one (a fork a repository cuts from the source's window is a new row with no floor, W19), and a whole log, where the entry and its restatement are one fact; `getWorkspaceState` takes a compaction's restatement as the nearest carrier when it is. The newest compaction is inside every bounded window and every entry below the floor was written before it, so a window cut in log order (the SQL backends' wake) whose newest compaction restated them reads what the whole log reads (a window whose newest compaction did not — KL-3468-a / KL-3468-c / KL-3468-d — reads as before): legs run in the session's tree, the release finds and releases it, the orphan doors see the record. A bounded store needs no change.
|
|
1383
|
+
- **Never vouched.** A restatement states only a fact of the log that later entries can add to (a release) but never overturn, and that does not depend on the view it was read through. A record is stated (an allocation is never undone; a release anywhere releases it). `unreadable` is stated whenever the reader vouches for it — on the whole log, and on a window that reads it from a restatement or from entries it holds (a restated `unreadable` stays unreadable through every later compaction, W5, W5b, W5w). The one `unreadable` a floor alone causes — a release above it whose allocation is below it, with no restatement — carries the window's floor, as an unvouched `none` does, and is not stated: stating it would turn the whole log, which reads the released record, unreadable for good (W21, measured before this rule: the whole log and every later window read `unreadable`). Not stating it costs this (KNOWN-LIMITS KL-3468-d): after a compaction on that window the next window reads `none` with its floor, so a leg that sends nothing about the workspace runs in the base checkout, while the whole log reads the record released and refuses a leg (measured by the review leg, two forms; new against the earlier cuts of this version, not against 7.67.1, which had no restatement). A `none` the reader cannot vouch for (a bounded window in which no compaction restates the session's checkout) is not restated: the next compaction leaves the key out, and the reader keeps answering `none` with `window: { floor }`. A vouched `none` is stated only where the session can no longer declare a checkout once the compaction lands — the one judgment `openSessionCheckout` refuses a declaration by (`checkoutDeclarationClosed`: a session that is no fork declares only at its birth; once its log holds a message or a compaction it has begun): a session that is no fork states it at every compaction (the compaction closes it as it lands, W22), a fork never does — a fork may allocate after a compaction (W20). A log whose compactions were written before 7.68.0 and whose record is below the floor therefore reads as before — the known residue, KNOWN-LIMITS KL-3468-a (reachable today by no bundled host: the file store loads the whole log; a session whose checkout was opened on 7.68.0 or later does not have it, unless the read a compaction makes of the record failed or the compaction's `details` is not an object, KL-3468-c). The residue reaches a fork copy too: a fork the repository copies from such a window is a new row with no floor and no restatement, and reads its source's checkout as a `none` it vouches for — carried, its rewind to the source's turn writes the tree's bytes into the base checkout (W19w); so does a fork copied along a path of any window that does not pass the window's newest compaction (W19p, read at the reader). Neither is introduced by this version (the review leg read W19p's shape the same on this version's tree before this round, `aea66b345`; on 7.67.1 a fork was carried whatever root its source had); server's SQL fork copies the whole stored log and does not reach them. Closes with #4328 (history rows that carry their own root).
|
|
1384
|
+
- **Known limit (KL-3468-c).** When the store read a compaction makes of the record fails (or the caller passes `details` that is not an object), that compaction states nothing; a window that starts at it reads `none` with its floor — the old reading — and a compaction made on such a window cannot state the record back (it cannot vouch). A failed read is not stated as `unreadable` (that would refuse a session with no checkout forever) and does not fail the compaction (carrying the facts must never be why a compaction cannot land). Reach: a host on a store that wakes bounded windows, with session checkouts, whose record read fails at a compaction (W12b pins it as it stands); or a host that calls `appendCompaction` with a `details` that is not an object (the engine's own passes do not; cli and server make no direct call).
|
|
1385
|
+
- **The out-of-run rewind (with #4276, same release).** `Runner.rewindFiles` reads the session's root through the same record reader, so on a compacted session woken from its window it previews and restores in the session's tree; the base checkout is not written (W14, both cuts). A session that never had a checkout rewinds at the deployment's root as before (W14c).
|
|
1386
|
+
- **A session that never declared a workspace (with #4280, same release).** Compacted and woken from its window, it reads an authoritative `none` (the compaction restated it): the undeclared leg and the `workspace: null` leg run at the deployment's root, the run's `rewindFilesTo` and `Runner.rewindFiles` rewind there, a declared leg is refused `config.session_workspace_changed` as on the whole log — and on a log whose compactions predate the restatement (`none` with its floor) every one of these answers the same: no reader of the session's root refuses for the floor (W14d, both logs). The one reader that does treat `none` with its floor as unknown is the fork carry (#4280, `recordedRootOf`): a fork woken from its window reads its source's root from the copied restatement and is carried (a rewind to a pre-fork point restores); a copy woken as a window whose compactions predate the restatement is not carried and that rewind is refused `rewind_snapshot.unresolvable`, nothing written (W14e, both logs) — the window is not carried when it holds no restatement to vouch with; a copy cut from such a window as a new row, with no floor, is (KL-3468-a, W19w).
|
|
1387
|
+
- **A fork with a tree of its own, across the floor (with #4280 Narrowing 4, same release).** The judgment "a fork that works in a tree of its own" (`forkHasOwnTree`) reads the fork's record through the record reader, so a window that cut the allocation entry still reads the fork's tree — from the restatement — and both rewinds refuse it there as on the whole log, neither checkout written (W16, both cuts). Measured before the restatement was read there: on such a window a fork that ran in the base checkout and then declared its tree read as holding none, and its run's `rewindFilesTo` to the earlier point answered `completed` and wrote the base checkout's bytes over the person's file in the tree (`tree-person` → `a`).
|
|
1388
|
+
- **A restatement that arrives through the store's import (session sync).** It passes the import door when well-formed and agrees with the session's allocation entry the same log holds, and is read like any copy of the session's own record (the same rule as every copy: its `sessionId` and its record's are the session's). Narrowing (open to veto @server): a restatement that says `none`, or holds another record (its release aside), while the same log holds that session's allocation entry is refused `invalid_session` — wherever either stands in the log, judged once the whole log is in, so an allocation on a sibling branch after the compaction is refused too (W8b, W8c); a restatement in a log with no allocation entry of its session is not judged (SECURITY-BACKLOG #4337). It does not let an import point the session at another tree: before a leg works in a tree of the record, or a rewind reads it, the tree's birth mark must name the record's allocation and the session's occupant (`workspace.checkout_missing` otherwise) — a forged restatement naming another session's tree is refused at the import door when the log holds the allocation it contradicts, and otherwise refused there, preview and restore alike, and that tree is not touched (W15b). An import onto an absent id is a new occupant: its record's trees are the source occupant's, and a leg there is refused the same way (W15a).
|
|
1389
|
+
- **Refusal added (open to veto @server @cli).** `Session.appendCompaction` refuses a caller whose `details` names `sessionCheckout` — `SessionError("invalid_entry")`, nothing written: the restatement states a stored record, and nothing but the engine writes one. A caller's `workspaceState` follows the existing rule for restated keys (the caller's explicit value wins). A non-object `details` is the caller's own shape and carries neither. A host whose own compaction copies the previous compaction's `details` into the next one must leave `sessionCheckout` out.
|
|
1390
|
+
- **Narrowing (open to veto @server @cli): a session that is no fork declares its workspace at its birth only — a compaction now closes it as a message does.** `openSessionCheckout` refused a declaration on a session that is no fork once its log held a message; it now also refuses one once its log holds a compaction (`config.session_workspace_changed`; the sentence is made true of both: `TaskSpec.workspace is declared on a session that already began without it (its log holds a turn or a compaction); a session's workspace is fixed at its birth. Start a new session to work in a declared workspace.` — it read `… already ran without it; a session's workspace is fixed by its first leg. …`; no match on the old words in server / cli / client-core / test, searched by the review leg). Before, the judgment read the loaded view, and a window holds no message when a host's own compaction names a non-message entry as its kept floor: a session that had run was allowed to declare a tree on such a window — measured: the declaring leg answered `completed` and minted a tree on 7.67.1's judgment and on the earlier cuts of this version (W23 now: refused, no tree, and the log passes its own import door). A window always holds the compaction its floor comes from, so the judgment reads the same on every view. Who is hurt: a host that compacts a session holding no message (a compaction of its own, over custom entries only) and then declares a workspace on it — refused now (W22). The engine's own compactions always keep a message (`compaction.ts`, the kept range is never emptied), so a Runner session reaches neither shape.
|
|
1391
|
+
- **Hard narrowing: `sessionCheckout` is the engine's key in a compaction's `details` (open to veto @server @cli @client-core @sdk, hosts that embed the engine directly).** A host that before 7.68.0 put its own value under that key name in a compaction's `details` — a string, or an object with no session id — reads, after the upgrade, as a restatement of the session's checkout in a shape that is not valid: the record reads `unreadable` and every leg of that session is refused `workspace.checkout_unavailable` (7.67.x read the record), and the import door refuses that log (`structurally invalid sessionCheckout restatement`); an object naming another session's id is skipped by the reader but refused by the door the same way (measured, review leg r2 D). cli and server mainline do not write this key (read from their code by the review leg).
|
|
1392
|
+
- **Read at the leaf it lands on.** `appendCompaction` now reads the leaf first and restates at it; the record's expected leaf is that leaf, so a write that lands while the restatements are read (another writer of the same session) makes the append a `SessionError("conflict")` — a compaction pass reads that as `leaf_moved` — instead of a compaction restating what the write replaced. Before, the leaf was read last. A store read that throws while the record is read leaves `sessionCheckout` unstated (it is never restated as `unreadable`, which only a log that holds an unreadable record states).
|
|
1393
|
+
- **Fork.** A fork's copied compactions restate the parent's checkout and are skipped (the fork holds no checkout of its own until its first declaring leg); a restatement that names the fork but holds another session's record is not a valid restatement (`unreadable`), never adopted. A fork's own compaction does not state its `none`: a fork compacted before it allocated, its allocation then left on a branch the leaf moved off (a later `resumeAt` above it, or a declaring `resumeAt` leg refused after it allocated), reads `none` with its floor on a window cut along the branch — and the orphan doors keep its tree — where a stated `none` would have read as vouched and let them remove it (W20, both ways). Narrowing (open to veto @server): a fork of a fork, woken from a window, reads its source's checkout as unknown and is not carried; a rewind to its source's turn is refused `rewind_snapshot.unresolvable`, nothing written (W20c).
|
|
1394
|
+
- **Session-log data (additive).** Two optional keys on a compaction entry's `details`: `sessionCheckout`, `workspaceState`. The import door (`validateEntriesForImport`) refuses a compaction whose key is present but malformed, or whose `sessionCheckout` contradicts the session's allocation entry the same log holds, in any order (`invalid_session`), and passes it absent. An engine before 7.68.0 reading a log written by 7.68.0 ignores both keys (it reads neither; its import door judges only the keys it names). Export surface: unchanged (no new root export). Wire frames: none.
|
|
1395
|
+
- **Follows the reference.** Claude Code 2.1.321 re-appends the session's metadata after every compaction so a tail load from the boundary still sees it (`reAppendSessionMetadataAfterCompaction`, cli321.js:285099, called at four sites on three lines of the compaction paths, :286980 (two) / :287590 / :287593; the re-appended set includes `worktree-state`, :285056-285060; `worktree-state` is `last-wins` in its tail-load table, :285031). Restating on the compaction entry itself is the same reading, landed in the one write that moves the floor.
|
|
1396
|
+
- **What a restatement does not cover (KNOWN-LIMITS KL-3468-b).** What a window leaves out above its floor: a window cut along the branch alone (`boundedTail`) does not load other branches, so a release written after the newest compaction and left on a branch the leaf was moved off reads as not released there — through the restated record (before this version that window read none with its floor; the SQL backends cut in log order and load the release); the workspace state is read along the branch, so a rewind to an entry between the floor and the newest compaction reads the task root.
|
|
1397
|
+
- **Not changed.** The record entry, its writers (`openSessionCheckout`, `releaseSessionCheckout`) and their codes; the orphan doors' and the open judge's readings of each answer; the file store (no floor); the other restated keys (`promptEpoch`, `announcedListings`, `gitAnnouncement`, `activeTools`, `thinkingLevel`, `model`). How a SQL backend picks its window (the newest compaction in the whole log, whatever branch) is its own rule: the restatement does not change which leaf a window holds; the checkout, a session-wide fact, reads the same from a compaction on any branch.
|
|
1398
|
+
- **Rules.** One carry for the session's standing facts at the one write that moves the floor, and one fold in the record reader (the restatement is one more copy), read through one read the carry shares, on every log (two more async frames on every leg's open: the rejection-depth pin moves 61 → 63); a restatement states only a fact of the log that later entries can add to but never overturn, read the same through any view; the one judgment "this session can no longer declare a checkout" (`checkoutDeclarationClosed`, extracted from `openSessionCheckout`) is read by the declaration's refusal and by the statement of `none`; one reading of "the view's, not the log's" (the reader's `window` note, on `none` and on the one `unreadable` a floor alone causes) is read by the statement of both. Added: the carry of two keys, one refusal at `appendCompaction`, three at the import door (a structurally invalid `workspaceState` or `sessionCheckout` restatement, and a restatement that contradicts the same log's allocation entry — the last judged in no order). Retired: none (the doors' `window` reading stays for the residue).
|
|
1399
|
+
- **Pins.** `test/backlog3468-checkout-record-across-floor.test.ts` — real Runner over a store that wakes every acquire from its bounded window (two cuts: the SQL backends' log-order cut and `boundedTail`): W1 compacted out of a run and woken, the undeclared leg, the declared leg, the orphan list, the record and the release; W2 three compactions in a row, the window reads the whole log after every wake; W3 a compaction inside a run; W4 an unvouched none is not restated as none; W4b an authoritative none is; W5 unreadable survives the floor (the leg is refused `workspace.checkout_unavailable`), W5b / W5w and through a host's compaction on a window that reads it from a restatement / from the entries it holds; W6 the residue; W7 fork; W8 the refusal and the import door, W8b the door's contradiction refusal, W8c (two variants) in no order — the allocation on a sibling branch after the compaction; W9 the workspace state; W10 the file store; W11 the copies fold (entry and restatement in one window, a release below the floor, one allocation stated two ways); W12 a store read that throws states nothing, W12b the known limit after it (KL-3468-c); W13 a write during the carry is a conflict; W14 (both cuts) the out-of-run rewind in the tree, W14c its control; W14d (both logs) a never-declared session: two legs, a declared leg and both rewinds; W14e (both logs) its fork's carry; W16 (both cuts) a fork with a checkout of its own: both rewinds refused on the whole log and on the window; W18 the fork judgment's own read fails (both rewinds and a plain leg refused `workspace.checkout_unavailable`), W18r the leg's row read fails; W19 (source in a tree / control) a fork a repository cuts from the source's window, W19w / W19p the known limit's two fork-copy forms (witnesses, KL-3468-a); W20 (two ways the leaf leaves a fork's allocation) a fork's `none` is not stated, the orphan doors keep its tree, W20c a fork of a fork; W21 a window's unreadable is not stated (the whole log keeps reading the released record); W22 (whole / window) a compaction on a session with no message states its none and closes it to a declaration; W23 a session that ran, a window with no message: the declaration is refused and the log passes its own import door; W15a / W15b the store-import path. Each mutation in the car's table reddens at least one cell.
|
|
1400
|
+
- **Consumers.** @server: the prerequisite of S-1950 (the session checkout over the SQL backends) — the two SQL stores need no new reading; a compaction they store carries `details.sessionCheckout` / `details.workspaceState` as payload; the window rule ([14067]) is separate. @cli: the file store reads the whole log, no change; its logs carry the two keys. @client-core: no wire change (a mirror of compaction details sees two more optional keys).
|
|
1401
|
+
|
|
1402
|
+
### The provider's own sentence (`providerMessage`) withholds what the credential vocabulary finds in its words as written, together with what the one pass finds in its one-line form — a key split by a zero-width or other format character, and a private-key body read by its line breaks, no longer ride the retry frame or the terminal refusal account in part (#4329; Security; shipped in 7.67.2; behavior change: narrowing — more of such a sentence renders `«redacted»`; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli @client-core)
|
|
1403
|
+
|
|
1404
|
+
- **Affected versions.** 7.67.0 and 7.67.1. The member is new in 7.67.0; 7.66.x carries no `providerMessage`, so this is a regression against the previous minor. The same change ships in 7.67.2; its notes there carry the readings against v7.67.1.
|
|
1405
|
+
- **Affected faces.** Every reader of the one function `providerMessageOf` (`src/brain/provider-message.ts`): `terminal.detail.providerMessage` on a `provider_rejected` refusal (a 4xx — the run's result, its `terminal` event and its replay); `providerMessage` on the retry status frame (`BrainStatus`, a retried 429 or other 4xx); a host Brain's `BrainStatus.providerMessage` and a host-minted terminal message's `providerMessage`, admitted through the same function. `terminal.message` / `errorMessage` were not affected: the engine reads the raw body there without folding it first.
|
|
1406
|
+
- **Trigger.** The provider's `error.message` holds a credential split by a format character (`\p{Cf}`: U+200B – U+200D, U+2060, U+FEFF, U+00AD, U+061C, U+200E / U+200F, the bidi overrides and isolates, the tag characters): the old order folded the sentence to one line before the pass read it, the fold turned the character into a space, and the vocabulary's format-free reading read only the part before it. Or the sentence holds a private-key body whose row reads line breaks (a PEM banner without its closing dashes followed by base64 lines, a `Private-Lines: <n>` body): the fold joined the lines and every body line was carried.
|
|
1407
|
+
- **Now.** The sentence is folded to one line and the one pass reads it, as before (`Authorization:\nabcdef…`, `Authorization: Negotiate\n<token>` stay masked); what the credential vocabulary finds in the sentence as written (as written and with its format characters removed) is folded the same way and taken out of the one-line form (since #4343, below, as values the vocabulary found — `remoteWords`' `found` argument, which no Cookie member's name spares — not beside the request's declared values). Both are found on the same unmasked words and their ranges merged, so neither hides the other. A sentence whose written reading the scanner cannot afford (more than 64 places a format character stood) is withheld whole. Then the 240-code-point bound. One function, one pass, one bound.
|
|
1408
|
+
- **One word more is withheld on one shape (#4332)** — `Authorization: Bearer<U+200B><token> retry later…` reads `Authorization: «redacted» later…`, as `terminal.message` reads it. **Known exception, host internals (#4335)** — a credential value found in the written words that is also a common piece of a URL or a path in the same sentence is taken out of it before the `human` tier reads it, which then shows the URL's host or the path's tail (never the credential). **Known, not changed here (#4333)** — a key split by a character the vocabulary reads as the end of a token (a line break, a tab, a control character, U+2028 / U+2029, a Unicode space) keeps its remainder on every face (#4343, below, closes it on this exit for a token behind a label or a scheme word whose separator the endpoint escaped, pin L4; a key with no label keeps it); a declared value split by a format character or a line break is not found by value; on `terminal.message` (JSON body) a `Private-Lines:` body and a PEM banner without closing dashes keep their body (KL-3769-b); a value that ends inside an engine tag is spelled differently once folded (closed on this exit by #4343, below, pin W7). **The bound and a second admission** — a cut can leave a tail the vocabulary reads as a credential once read again; the source comment says so. **The same order elsewhere (#4334)** — other mints fold untrusted words before a later throat masks them (A2A error words, host seat conditions, hook and extension reasons, memory announcements, rule and list text).
|
|
1409
|
+
- **On this tree (7.68.0 candidate), measured:** `terminal.message` over a JSON body holding a fenced PEM followed by a line of plain words takes the escaped line break and the next word with the PEM's mark (`bad key: «redacted» later, …`); the carried `providerMessage` keeps the line (`bad key: «redacted:private-key» retry later, …`). The pin's raw-body sentinel names that shape beside the #4332 one.
|
|
1410
|
+
- **Legibility.** A sentence whose written reading has more than 64 places a format character stood is withheld whole, as `terminal.message` withholds it.
|
|
1411
|
+
- **Rules.** Passes over the sentence: 1 → 1 as this section left it (#4343, below, reads it three ways — the raw body, the sentence as written, the folded sentence's echo pass; KL-4343-c); what that pass takes out: the declared values → the declared values and the parts the vocabulary finds in the words as written; functions: 1 plus one module-internal reader, `egressCredentialParts` (`src/core/untrusted-egress.ts`; not root-exported); bounds: 1, after the pass. Special cases: 0.
|
|
1412
|
+
- **Pins.** `test/backlog4329-provider-message-mask-order.test.ts` (78 shapes through the exits — a real Runner, and the host-minted terminal through `terminalCauseOf`; ① – ⑦ as in the 7.67.2 notes). Re-pinned: `test/backlog1824-2130-provider-message.test.ts` S7 / H3 (titles); `test/fixtures/backlog3826-inline-untrusted-authors.json` (the function's one row, its bound call over the folded sentence).
|
|
1413
|
+
- **Consumers.** No code change asked. @server @cli @client-core: `providerMessage` no longer carries the parts named above.
|
|
1414
|
+
|
|
1415
|
+
### A remote's words lose every declared value they hold, also where two values overlap — the value matcher behind `remoteWords`, `scrubRequestEcho` and `eraseSentCredentials` finds a value that starts inside another's hit (#4329; Security; shipped in 7.67.2; behavior change: narrowing — where two values overlap in the words, the later one's tail renders `«redacted»` too; widening: none; hard-breaking: none; closed sets: none; exports: none; wire: none) (@server @cli)
|
|
1416
|
+
|
|
1417
|
+
- **Before.** The matcher took the longest value at a place and resumed after that hit, so a value that started inside it was never found (`abcdefgh-ijklmnop` with `abcdefgh-ij` and `h-ijklmnop` to take out read `«redacted»klmnop`).
|
|
1418
|
+
- **Now.** The search resumes one character after each hit's start; overlapping hits are merged into one mark. Every face that mints a remote's words through these three functions withholds more where values overlap and nothing less (a model brain's failure text and `providerMessage`, an MCP server's failure and listing, an A2A peer's card and error words, the memory distiller seat's failure).
|
|
1419
|
+
- **Pins.** `test/backlog4329-provider-message-mask-order.test.ts` ⑦ (V1 every shape against the old matcher; V2 two overlap shapes through `scrubRequestEcho` and `eraseSentCredentials`; V3 a control on the distiller's real seat).
|
|
1420
|
+
|
|
1421
|
+
### undici 7.30.0 → 7.30.2 (exact): the advisories undici 7.30.1 closes are closed where THIS package's undici copy carries the layer they sit in — the connection when its Agent is the process dispatcher, the decoding only when a host passes its `fetch` as `fetchImpl` (the engine never calls that fetch) — on Node 24 a plain package-root process sends through the runtime's own bundled undici instead, which this pin does not reach (#4350; dependency bump; behavior change: four, below, only where this package's copy carries the request, two of them read from the diff only; narrowing: one, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none) (@server @cli @client-core @test)
|
|
1422
|
+
|
|
1423
|
+
- **What changed.** `dependencies.undici` `7.30.0` → `7.30.2`, exact; the lockfile entry carries the registry's integrity (checked against the packed tarball); the package's `dependencies` name one copy (the repository's `undici6` / `undici8` aliases are dev-only). The package uses undici for the transport seat only (`src/brain/transport.ts:1-3`: the Dispatcher base class, the slot reader, the two handler converters, all byte-identical in 7.30.2). It never calls its copy's `fetch`, composes no interceptor, uses no cache store, EventSource or WebSocket, and builds no Agent.
|
|
1424
|
+
- **Which copy carries a request.** Every engine request (model calls through the engine dispatcher; Node's own `fetch`, which WebFetch, MCP over HTTP and A2A use) is CONNECTED by the Agent in the process slot `undici.globalDispatcher.1` (TLS, HTTP/1.1) — a caller speaking Dispatcher API v2 (an undici 8 `fetch` passed as `fetchImpl`) by the one in `undici.globalDispatcher.2` when a host filled it (`src/brain/transport.ts` `hostDispatcher`) — and DECODED by the caller's fetch layer: the runtime's own `fetch` unless a host passes `fetchImpl`. Measured on Node 24.21.0 (bundles undici 7.29.1) and 24.2.0 (7.10.0): after a fresh package-root import the slot holds the runtime's Agent. The root graph's `node:http` import loads the runtime's undici before this package's copy evaluates, and undici fills only an empty slot. This package's Agent carries requests only when a host installs it (`setGlobalDispatcher(new Agent())` from this copy, which is also the copy a host pinning the same version shares), or on a runtime whose `node:http` does not load undici (Node 20.18.1 per KL-1121c, an earlier reading, not re-measured here).
|
|
1425
|
+
- **The eight advisories of 7.30.1, by reach.**
|
|
1426
|
+
- GHSA-qrxm-fqhp-jwvm (a TLS 1.2 peer renegotiates to a CA-trusted certificate for another name; the process exits on an assertion). Reached by any TLS endpoint the engine is pointed at: a model base URL, a WebFetch URL, an MCP or A2A server. Measured with this package's Agent in the slot: on 7.30.0 the process died on a model call and on a plain fetch; on 7.30.2 both end in `[network] fetch failed (ERR_TLS_CERT_ALTNAME_INVALID: …)` and the process lives. With the runtime's Agent in the slot (a plain package-root process on Node 24.21.0, runtime undici 7.29.1) the process died on both pins.
|
|
1427
|
+
- GHSA-2m5v-pch6-74hx (Agent drops a function-valued `tls` option). Not reached by the package, which builds no Agent. Reached through a host that builds this copy's Agent with `tls.checkServerIdentity`. Measured: on 7.30.0 the callback never ran and both engine calls succeeded; on 7.30.2 it runs, the calls end in `certificate pin mismatch`, and the endpoint receives no request.
|
|
1428
|
+
- GHSA-xjj6-f33q-929c (more than five content codings, then an early close: an unhandled stream error exits the process). In the fetch layer, never this package's copy unless a host passes its `fetch` as `fetchImpl`. Measured: with this package's Agent in the slot, the runtime's fetch (7.29.1) died on both pins. With this copy's `fetch` as `fetchImpl` the process died on 7.30.0 and lived on 7.30.2.
|
|
1429
|
+
- GHSA-4w7w-5xrc-9h26 (cache + redirect interceptors) and GHSA-55wc-3pg8-jfw6 (decompress interceptor): not reached. The package composes nothing; the one `compose` in `src/` is the seat's pass-through (`transport.ts:131-139`).
|
|
1430
|
+
- GHSA-4qv8-2p9j-2685 (EventSource): not reached. `src/` has no EventSource; the MCP client parses SSE with `eventsource-parser`, and the `eventsource` package serves only an SSE transport `src/` does not import.
|
|
1431
|
+
- GHSA-9fm3-97px-rjf6 (WebSocket PONG): not reached; `src/` has no WebSocket client.
|
|
1432
|
+
- GHSA-f72v-6w45-q472 (SRI token case): not reached. No request in `src/` carries `integrity`, and this copy's fetch is never called.
|
|
1433
|
+
- 7.30.2 closes six more (GHSA-mqg3-3v4r-cm32, -7xg3-fphp-28hg, -5rm8-gv32-r954, -9v6h-8rgp-m69f, -3g9m-rv3r-w5w2, -r4pc-4jcp-7pqc: cache, redirect interceptor, cache replay, SRI whitespace). None is reached, for the same reasons.
|
|
1434
|
+
- **Behavior changes, only where this package's copy carries the request.** ① A renegotiation to a mismatched certificate is a call failure (`network`; read as transport and retried on the shared ladder, per `classifyFetchThrow` / `readConnectFailure`, not measured with retries on) where it was a process exit. ② A function-valued `tls` option on an Agent built from this copy is applied. ③ The HTTP/1.1 client no longer sets each socket's type of service to 0 when the runtime has `socket.setTypeOfService` (Node 24.21 has it), and a throw from it no longer fails the request (read from the diff, not measured). Also from the diff, not measured: the TLS session cache keeps a new session when it evicts a collected one. Unchanged in the diff: error codes (`lib/core/errors.js`), timers, `lib/global.js`, the proxy agents.
|
|
1435
|
+
- **What `npm audit` says (measured, registry advisory database, empty user config).** The published 7.67.1 installed into an empty project: 1 high (undici 7.30.0, the eight above) and 1 moderate (`@sema-agent/core`, through undici). npm's suggested fix is `@sema-agent/core@7.32.1` (`isSemVerMajor: true`), so `npm audit fix --force` DOWNGRADES the package to 7.32.1, the last release before undici entered `dependencies`. The candidate built from this change (`npm pack`, installed into an empty project) resolves `undici@7.30.2`, and `npm audit` reports 0 vulnerabilities and suggests nothing.
|
|
1436
|
+
- **Narrowing (open to veto @server).** A host Agent from this copy whose `tls.checkServerIdentity` refuses a certificate now refuses the engine's request; on 7.30.0 that callback was dropped and the request went through.
|
|
1437
|
+
- **Not changed: the residual (KL-4350-a).** A plain Node 24 process that imports the package root sends through the runtime's undici. Each of the two process exits stays open there while the runtime's bundled undici is in that advisory's own range. GHSA-qrxm: below 7.30.1 (6.x below 6.29.1, 8.x below 8.12.2). GHSA-xjj6: from 7.18.0, below 7.30.1 (6.23.0 to below 6.29.1, 8.x below 8.12.2). Measured on Node 24.21.0, which bundles 7.29.1 (in both ranges): both exit the process on either pin. Measured on Node 24.2.0 (darwin), which bundles 7.10.0 (in GHSA-qrxm's range only): the renegotiation exits the process, and the six-codings close ends the call in `[network] mid-stream read failed: terminated (UND_ERR_SOCKET: other side closed)` with the process alive. Remedy: a Node release that bundles undici 7.30.1 or later, or a host that installs an Agent from a fixed undici (covers the connection layer, GHSA-qrxm) and passes a fixed undici's `fetch` as `fetchImpl` (covers the fetch layer for the brain, or the WebFetch tool, it is given to; MCP and A2A use the global `fetch`). The engine does not install a dispatcher of its own (`transport.ts:16-17`).
|
|
1438
|
+
- **Consumers.** @server: pins `undici` `7.30.0` exact (`sema-server` origin/main `package.json`, read on 11-13; its lockfile holds one copy, `node_modules/undici` 7.30.0). After this bump an exact 7.30.0 at the server's root cannot share this package's 7.30.2, so npm keeps the server's copy at the top and nests this package's under it (npm's resolution rule, read, not installed here); a dispatcher the server builds from its own `undici` stays 7.30.0 until the server moves too. @cli: `undici` `^8.5.0` (`sema-cli` origin/main `package.json`, read on 11-13; lockfile top-level 8.7.0; the same advisories' 8.x fix is 8.12.2). @client-core: no direct undici. @test: black box. On a host that installs this package's Agent, a TLS 1.2 model endpoint that renegotiates to a CA-trusted certificate for another name fails the call and the process lives. On a plain Node 24 host it still exits the process (KL-4350-a).
|
|
1439
|
+
- **Pins.** `test/backlog4350-undici-advisory-reach.test.ts`: ① the slot owner after a package-root import; ② a host-installed Agent is the pinned copy, at 7.30.1 or later; ③ ④ renegotiation through a model call and a plain fetch; ⑤ the runtime-owned residual; ⑦ ⑧ six codings through the runtime's and the package's fetch; ⑥ the host's `tls` pin. `test/backlog1181-spawn-site-env.test.ts` reads the reviewed undici spec from the review row (`scripts/verify-spawn-sites.mjs`, re-read at 7.30.2: starts no process).
|
|
1440
|
+
- **Rules.** None added, none retired. The spawn-review test reads the reviewed version from its row: one copy of the version, not two.
|
|
1441
|
+
|
|
1442
|
+
### The README quick start (both languages), the SDK getting-started page and the zero-config table now say what the package does — a result's ending is `terminal.kind`, the key goes on the brain beside the model's own address, the SQL example declares its pool, the engine's own file-write tools are asked about by default, and the memory row spells `scopes` (documentation only)
|
|
1443
|
+
|
|
1444
|
+
- **Before.** The README quick start printed `res.status`, which is `undefined` (a `TaskResult` has no such member; the ending is `terminal.kind`: `completed` | `failed` | `blocked` | `paused`), and the highlights listed a status set (`timeout`, `suspended`, `needs_review` among it) that `terminal.kind` does not take (the five-word `TaskStatus` the root exports has no `timeout` and is not a member of the result either; `terminalProjection(res.terminal).status` derives it). It had no key; the obvious way to add one — the key on `createOpenAIBrain`, the address only on the model entry — fails the task with `route.credential_mismatch`. The README and `docs/sdk/01-getting-started.md` required Node >= 20 where `engines` says `>=20.18.1`. The "complete, runnable integration" in `docs/sdk/01-getting-started.md` used a `readonlyPgPool` it never declared. The zero-config table in `docs/sdk/10-extension-points.md` said tool review is **None** by default; the engine's own `Write` is asked about under the default permission mode and, with no approver wired, the ask is refused and the file is not written. The same table's memory row still wrote the retired singular `memory: { scope }`, which a task start refuses with `config.memory_scope_spelling`.
|
|
1445
|
+
- **Now.** The quick start (`README.md`, `README.zh-CN.md`) is one minimal example that was run through the packed package: project setup with `type=module`, a seven-member `Model`, the key and the address given to `createOpenAIBrain` together (`apiKey: process.env.MODEL_API_KEY`, absent = no credential header), and `res.terminal.kind` plus `res.result` printed. The highlights line names the four `terminal.kind` words and says the result has no `status` member and the older five-word status is derived on demand (the derivation stays told in `docs/sdk/02-runner-and-tasks.md`; the README names no export for it, so no export's documentation tier moves). The quick start's run line says a `MODEL_API_KEY` already in the shell wins over one loaded with `--env-file`. Both Node floors read `>= 20.18.1`. The SQL example declares its pool from `pg` and says what it needs (`pg`, a read-only role's connection string). The zero-config row says a tool passed in `tools` runs without a question unless it declares `permissionSeat: "passthrough"`, that the engine's file-write tools ask by default and are refused with no approver, and names `onAsk` and `permissionMode: "acceptEdits"` as the ways to answer; the memory row spells `memory: { scopes: ["…"] }`.
|
|
1446
|
+
- **Not changed.** Source, wire, the package's behaviour and every export. The rest of both READMEs and of the two SDK pages is as it was.
|
|
1447
|
+
- **Rules.** Retired: none. Added: none.
|
|
1448
|
+
- **Consumers.** None: documentation text only.
|
|
1449
|
+
- **Wire.** None.
|
|
1450
|
+
|
|
1451
|
+
### `providerMessage` is no weaker than the same response's `terminal.message`: what the raw-body reading withholds inside the sentence's literal, the sentence withholds too — a token split inside by a line break, a tab or an escaped Unicode space, a value the body labels elsewhere, no longer rides the retry frame or the terminal refusal account in part (#4343; Security; shipped in 7.67.3; behavior change: narrowing — more of such a sentence renders `«redacted»`, see Legibility; widening: a value read inside an engine tag now leaves the rest of the sentence readable, see Now; hard-breaking: none; closed sets: none; exports: none from the package root; wire: none) (@server @cli @client-core)
|
|
1452
|
+
|
|
1453
|
+
- **Affected versions.** 7.67.0 – 7.67.2 and this line before the change. 7.66.x carries no `providerMessage`; there the same text reached only `terminal.message`, which read it clean.
|
|
1454
|
+
- **Trigger.** The provider's `error.message` holds a credential whose middle is a character the endpoint's JSON writer wrote as an escape and the vocabulary reads as the end of a token once it is parsed — a line break (LF, CRLF), a tab, VT, FF (every writer escapes those), U+2028 / U+2029 (Go too), a no-break space or U+3000 (an ASCII-only writer: Python's `ensure_ascii`, PHP's default) — or a character the writer escaped that the parsed sentence's reading needs as it is (a no-break space before a label's `=`, a zero-width character inside `Bearer`, `/` written `\/` behind a `Token` run); behind a scheme word (`Authorization: Bearer|Basic|Negotiate|<any scheme>`, a bare `Bearer`, `Token`) or a key label (`x-api-key:`, `api_key=`, `api_key:`, `api-key=`, `password=`). Or a value the vocabulary finds under a label elsewhere in the body (`"x-api-key": "<value>"` among echoed headers) stands again in the sentence and the body also holds a Cookie member or the request's declared key, which run the raw-body reading's echo pass: that pass takes the value out wherever it stands. Or a later pass of the vocabulary reads a value over the words an earlier mark set apart (`AKIA…password=<token>`). `terminal.message` reads the raw body and withholds it; `providerMessage` read only the parsed sentence and carried the rest.
|
|
1455
|
+
- **Now.** The one reader of the envelope (`statedSentence`, `src/brain/errors.ts`) states the sentence with where each of its characters stands in the body — the literal found by a walk that reads the body by `JSON.parse`'s rules (trimmed as the envelope reader trims it, member names decoded, the last of a repeated member) and checked by decoding (a literal that does not decode to the sentence withholds the sentence whole). What the raw-body reading withholds (`remoteWordsSpans`, `src/core/protocol-http-auth.ts`, beside `remoteWords`, which `terminal.message` is: its echo pass's spans — every Cookie member's value, the declared values, the values the vocabulary found anywhere in the body — and what its vocabulary pass then withholds in those marked words, read off the renderer's own account, `egressCredentialReading`, `src/core/untrusted-egress.ts`: every hit of every pass, each less the label its rendering keeps) is read back through the marks and mapped back through the literal's escapes onto the sentence, then taken out with the parts read in the parsed sentence and the declared values. On this line the marks' places are read back off `renderWithheld`'s own output (the spans merged as its regions, each member found where the verbatim text before it ends); a region it widened over a mark the body already held is not read back, and the whole body is withheld. The spans are checked against the reading's own output: where the text outside them does not stand in it as written, they missed a removal and the answer is the whole body. The pass reads the sentence folded only (`foldToOneLine`, `src/core/untrusted-text.ts`); engine tags are neutralized once, at the bound (`inlineUntrusted`, byte for byte as before), so every occurrence of every part is taken out: `password=<system-reminder attr=x>` reads `password=«redacted» attr=x>`. A host Brain's sentence (`admittedProviderMessage`) comes with no body and is read as written, as before. The shared matcher (`remoteWords`): a value the vocabulary found — here or (the new `found` argument) in another spelling of the words — is taken out wherever it stands, a Cookie member's name included; only a declared value spares that name.
|
|
1456
|
+
- **Measured.** As on 7.67.3 (its CHANGELOG section): 0 `providerMessage` cells carry a piece the same response's `terminal.message` withholds, against 1 690 before; no regression; server's 189-form cut table clean on its acceptance set. On this line's marks, the pin's exact cells write the vocabulary's own member (`«redacted:aws-key»password=«redacted»`).
|
|
1457
|
+
- **Legibility (KNOWN-LIMITS KL-4343-a, KL-4343-b).** A Unicode space right behind a credential that the endpoint wrote as an escape (U+3000 or a no-break space from an ASCII-only writer; U+2028 / U+2029 from Go or an ASCII-only writer) joins the words after it, up to the next ASCII white space, quote, `;`, `|` or `&`, to the credential, as the raw-body reading reads them (`Authorization: Bearer <token> 请检查您的密钥是否正确` from an ASCII-only writer → `Authorization: «redacted»`); a writer that left it as it is keeps the clause. A `Cookie:` / `Set-Cookie:` written as words in a string member before the sentence takes the sentence with it on both readings; a record member does not. Where the spans do not account for the raw-body reading's output, or a region was widened over a mark the body held, the sentence reads `«redacted»` whole (seeded texts: 0.36 %, runs of labels with nothing between them; 0 of the 135 envelopes the test suites hold).
|
|
1458
|
+
- **Performance.** As on 7.67.3 (its section has the measured cells): about four times the previous reading on 64 KiB bodies full of short distinct credential values, about 1 s on a host at load 9 – 12 and 3.06 s on another (KNOWN-LIMITS KL-4343-c), linear in the number of hits; the body read is capped at 64 KiB and a capped body states no sentence. Once per attempt that reads a body.
|
|
1459
|
+
- **Host seat (`admittedProviderMessage`).** Read as written only, by the same pass, so the pass's changes reach it: a value split by a zero-width character has its part after the split withheld where it stands again (a one-character part in every word that holds it, KL-4343-d); a value read inside an engine tag; no sentence carries a piece of a credential the earlier reading withheld (pin ⑪, over its five sets of sentences).
|
|
1460
|
+
- **Known, not changed here.** A token split by a plain space, a vendor-prefixed key with no label split by a line break, a scheme word itself cut by a line break (#4333); a private-key body whose banner was cut away (KL-3769-b, both exits); a value labelled elsewhere in a body with no Cookie member and no declared value (both readings carry it).
|
|
1461
|
+
- **Rules.** Readers of the envelope: 1 (`statedEnvelope`); `statedSentence` reads through it and walks the body once more to place the sentence's literal, checked against it by decoding (a literal that does not decode to the sentence is withheld with it). The raw-body reading's plan: 1, shared by `remoteWords` and `remoteWordsSpans` (`echoTaken`). What the vocabulary pass withheld, read off the body: the scanner's spans → the renderer's own account, which the echo pass's found values read too (`vocabularyValues` reads it). The sentence's as-written parts: the scanner's hits as before (a hit inside a larger one included) with the renderer's account. Folds before the pass: 1, the fold alone. Guessed writer spellings: 0. The Cookie-name exemption: declared and found values → declared values only.
|
|
1462
|
+
- **Pins.** `test/backlog4343-provider-message-not-weaker.test.ts` (as on 7.67.3, on this line's marks; E2 compares `inlineUntrusted` with this line's spelling before the change; one more cell for the marks' read-back; ⑪ against this line's reading before the change, spelled from exported pieces and equal to it on three plain sentences in the pin; the 8 000-input equality check is the car's reading, not a pin cell). Re-pinned: `test/backlog4329-provider-message-mask-order.test.ts` W7, `test/fixtures/backlog3826-inline-untrusted-authors.json`.
|
|
1463
|
+
- **Consumers.** No code change asked. @server @cli @client-core: `providerMessage` no longer carries the parts above; `@sema-agent/primitives` is unchanged by this section (#3970, above, changes it in this release).
|
|
1464
|
+
|
|
1465
|
+
### A declared A2A peer the run's tool valve can leave nothing of is sent no card request — the whole-face switch (`excludeAllTools`) no longer fetches the card of a peer whose tools it then unmounts; the dial now asks the reading the MCP dial already asked (#4357; behavior change: the dial of an A2A peer; narrowing: one on the network side, below, open to veto @server; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none) (@server)
|
|
1466
|
+
|
|
1467
|
+
- **Before.** A run with `excludeAllTools: true` that declared an A2A peer still sent the peer `GET /.well-known/agent-card.json` (and `GET /.well-known/agent.json` when that path did not answer) and then unmounted the peer's tools: "this leg mounts nothing" did not hold on the network. The MCP half already read the valve before it connected (`valveLeavesPeer` through the dial seat in `src/core/runner/prepare-protocol-tools.ts`); the A2A half did not. Two list forms that reach none of a peer (an `excludeTools` naming the whole peer, `a2a__peer__*`; an `allowTools` naming nothing of it) dialed the peer the same way.
|
|
1468
|
+
- **Now.** `materializeLegA2aTools` takes a trailing optional `dialable` predicate — an engine-internal parameter, not index-exported and not a host seat: only the Runner passes it, with a pure predicate that answers a boolean and does not throw, so it has none of the guard the MCP materialization's host-suppliable `dialable` has — and the run binds it to the reading the MCP dial already asks, the one server-level reading of the valve, `serverReached`, under the A2A prefix: a peer it answers `false` for stays declared and is not contacted — no card request, no status row, no tool, no warning; the handle's `refresh` holds the same answer, as the MCP refresh does for a server it did not dial (the all-peers form does not list the peer; naming it reads `failed` with "not dialed for this task" and sends nothing, since this status vocabulary has no `not_connected`; no run verb calls an A2A refresh today — read). The run's own tool lists are judged against the DECLARED peers for both protocols (`declaredPeers: { mcp: lockedPreflight.mcp, a2a: lockedPreflight.a2a }` in `src/core/runner/prepare-task.ts`; the A2A side read the dial's status rows, so a peer not dialed would have been told "no A2A peer named … is declared for this task" by a list that names it).
|
|
1469
|
+
- **Narrowing (open to veto @server).** An A2A peer is no longer asked for its card on a run whose valve's per-peer reading leaves nothing of it — the whole-face switch, an `allowTools` that names nothing of the peer, an `excludeTools` that names the whole peer; a pair of lists that cancel each other tool by tool is not that reading (below). A deployment that watched the peer's access log for a card request on such a run (a warm-up, a liveness signal) sees none. The peer's `statuses` row is absent for it, as an undialed MCP server's is, and so is its "peer unavailable" warning (`a2a.peer_unavailable`, phase `a2a`): a peer that would have been unreachable is not tried, so a deny list naming it hides the outage too. The one reader of a peer's status rows left in this repository is the handle's own `refresh` (read); the list judge above reads the declarations.
|
|
1470
|
+
- **Where the two protocols still differ (read, not changed).** With `excludeTools: ["mcp__srv__*"]` the MCP server is dialed all the same: the cross-server resource tools (`ListMcpResourcesTool` and its two siblings) stay mounted and aggregate what every connected server advertises, so the MCP dial reads `resourcesReached || serverReached(…)`. A2A has no such face, so under the same form the A2A peer is not contacted. Excluding the three resource tools too makes the MCP server undialed as well. The reading is per PEER for both protocols, so an `allowTools` and an `excludeTools` that name the same single tool of a peer leave no tool of it yet still dial it (the valve cannot know a peer's tools before the dial); the A2A peer and the MCP server get the same answer there.
|
|
1471
|
+
- **Not changed.** A run without the valve fetches the card and mounts the peer's tools; the name-collision check still runs over every declared peer, dialed or not; an A2A peer that answers, and one that was declared and never answered, read the same list judgment as before (no advisory in either state, on base and on head); a name no declaration answers to (the key absent, or an empty list) is still told as undeclared; `A2aPeerStatus`, `MaterializedA2a` and `materializeA2aTools` (the library entrance, which has no valve) keep their types; the `MaterializedA2a` JSDoc that ships in the declarations now says `statuses` holds one entry per DIALED peer and `refresh` contacts no peer the dial kept out (for a library caller, every declared peer, as before).
|
|
1472
|
+
- **Rules.** Readings of "may this peer be dialed" 2 → 1 (`serverReached`, asked by both protocols); readings of "which peers did the task declare" 2 → 1 (the declarations; the A2A side no longer reads the dial's attempt records). Added: none; retired: none as a rule (two readings folded).
|
|
1473
|
+
- **Consumers.** @server: a run under `excludeAllTools` stops contacting declared A2A peers (the access log of a peer that only chat lanes name will go quiet); nothing to change. @cli @client-core: none (no frame, key or word moves). @test: the black-box read is the peer's request log under the switch.
|
|
1474
|
+
- **Pins.** `test/backlog4357-a2a-unmount-valve.test.ts` — S1 real `Runner`, loopback A2A peer: under `excludeAllTools` zero requests and no `a2a__` tool offered (base: one card request) · S2 control: the same declaration without the switch fetches the card and offers `a2a__helper__summarize` · S3/S4 the MCP half under the switch: zero requests, and the dial without it (unchanged) · F1–F6 the valve's faces for both protocols (no valve; the switch; a list naming the whole peer; the same beside the three resource tools; an allow naming the peer; an allow reaching nothing of it) · F7 an allow and a deny naming the same one tool: the A2A peer and the MCP server get the same answer · L1–L10 a list naming a declared peer raises no "not declared" advisory, per protocol, and for an A2A peer in each state (answers / never answers / kept from the dial) under a whole-peer list and a one-tool list; a task that declares no peer (the key absent, an empty list) is still told · N1 a peer declared `Team.Lead` is the peer `a2a__Team_Lead__*` names, dialed or not · H1–H4 the materialization's own handle (the predicate answers no: nothing sent at the dial, and a refresh that names the peer says it was not dialed and sends nothing; beside a dialed peer the all-peers refresh walks only the dialed one; control; the predicate is asked before any request). Re-pinned: `test/backlog1623-tool-selector.test.ts` ③c, the unreachable-peer cell — the peer's warning is told on the allow list (the peer is dialed) and no longer on the deny list that names the whole peer (it is not).
|
|
1475
|
+
|
|
1476
|
+
### `excludeAllTools` is the leg's own allowlist naming nothing: a leg under it keeps the tools it relies on (a run with an `outputSchema` delivers `structuredOutput`, a goal iteration keeps `DeclareDone`), and `runWithVerification`'s verifier carries it and is handed its verdict tool alone (#4356; behavior change; narrowing: one, open to veto @server @cli @client-core; widening: four; hard-breaking: none; closed sets: none; exports: none added, none removed; types: none) (@server @cli @client-core @test)
|
|
1477
|
+
|
|
1478
|
+
- **Rule.** One: the leg's own lists never decide a tool the leg relies on (the catalog's protocol column: a delegated child's `ReportBlocked`, a teammate's `SendMessage` / scheduler / task-list tools as its spawner allows them, a goal iteration's `DeclareDone`, an observer's `ObserverReport`, a structured run's `StructuredOutput`), and the whole-face switch is one of those lists — an allowlist that names nothing. The switch is read once, into the run's own selector layer — after the run's own lists are read as written, so a malformed `allowTools` beside the switch is refused exactly as it is without it — so the selector itself — the roster valve, `ToolExecuteContext.toolSelector`, the ceiling a child inherits — leaves exactly what `allowTools: []` leaves. The reference behaves the same way: `--tools ""` narrows the built-in set and the structured-output tool is added after it, and a verifying hook agent is handed its parent's tools plus its verdict tool.
|
|
1479
|
+
- **Before.** The valve read the switch as "every tool unmounts, the relied-on ones included". A run with a valid `outputSchema` under it completed with no `structuredOutput` and no error (a call of `StructuredOutput` read `Tool StructuredOutput not found.`); a `runGoal` iteration under it lost `DeclareDone` and ran out its `maxIterations`; a `runCascade` under it climbed every rung and ended `exhausted`. The verifier leg of `runWithVerification` copied the implementation's `allowTools` / `excludeTools` / `deferTools` / `alwaysLoadTools` / `restoreGatedTools` but not the switch: on a Runner with an execution environment it was handed `ReportFindings`, `StructuredOutput`, `Read`, `Grep`, `Glob`, `RepoMap`, the read-only `Bash` and `ReadToolResult`, and read files the leg it judged could not; without one it was handed `ReportFindings`, `StructuredOutput` and `ReadToolResult`.
|
|
1480
|
+
- **Now.** Under the switch a structured run is handed `StructuredOutput` alone (the request's `tools` and the `wiring_manifest` roster both name it) and returns what it submitted; a goal iteration is handed `DeclareDone` alone; a cascade's first rung that submits passes. The verifier's spec carries the switch beside the five lists: it is handed `StructuredOutput` alone, a call of `Read` reads `Tool Read not found.`, and its verdict lands. A top-level leg that relies on nothing (no schema, not a goal iteration) still sends its request with no `tools` key and reads `entries: []` on its roster; a delegated leg relies on `ReportBlocked`, so under the switch it keeps that tool (Widening below); a server's tools are still neither dialed nor mounted; an own `excludeTools` that names a relied-on tool is still refused (`config.tool_face_control`, `detail.reason: "protocol_denied"`), with or without the switch; an invalid schema is still refused (`config.invalid_output_schema`).
|
|
1481
|
+
- **Narrowing (open to veto @server @cli @client-core).** A verifier of a leg under `excludeAllTools` no longer mounts hands, scaffolds or the tools `VerifyConfig.verifierTools` names: it judges from its objective and the evidence alone. A deployment that wants a verifier with hands runs the implementation leg without the switch (an `allowTools` narrowing rides to the verifier as before).
|
|
1482
|
+
- **Widening (four).** Every leg kind keeps its relied-on tools under the switch. (1) A structured run mounts `StructuredOutput`. (2) A goal iteration mounts `DeclareDone`. (3) A delegated leg keeps `ReportBlocked` (not with `enableBlockedReport: false`): its request and roster go from empty to `["ReportBlocked"]`, and a call of it ends the leg `blocked` (before: no `tools` key, `completed`). Three lanes hand a delegated leg the switch: a workflow `ctx.agent(spec)` whose spec carries it (a workflow marks every agent leg delegated), a host that starts a leg as delegated, and the workflow governance baseline's `base` (a `Partial<TaskSpec>` folded into every agent a script starts — measured on the Workflow tool's own launch with the repository's vm script runner). (4) A workflow agent with the switch and a `{ schema }`: before, it threw `workflow.agent_schema`; now it delivers its result (request `["ReportBlocked", "StructuredOutput"]`), or ends `blocked` when the model calls `ReportBlocked`. An observer keeps `ObserverReport` and a teammate the team tools its spawner allows (pinned on the selector; no real lane measured — an observer is spawned only by the Agent tool, which a switched leg does not mount).
|
|
1483
|
+
- **Not changed.** What `allowTools`, `excludeTools`, `deferTools`, `alwaysLoadTools` and `restoreGatedTools` do; which tools a leg relies on; the denylist refusal and its sentence; MCP dialing under the switch (A2A dialing is not this section's: #4357, above, sends no card request to a peer the valve leaves nothing of, which under the switch is every peer); the verifier's prompt (it still names running checks — a verifier with no hands may answer `PARTIAL`, `docs/KNOWN-LIMITS.md` KL-4356-a); when the verifier and fix legs read the caller's tool-face controls (at their own dispatch, as the lists always were — a host that changes its own spec between legs, the switch included, gets the changed value, KL-4356-c); a value of the switch other than `true` (`false`, `"true"`, `1`, `null`, `{}`) still reads as off, with no refusal and no notice (#4368).
|
|
1484
|
+
- **Rules.** Retired: two — the written exception "the whole-face switch unmounts the relied-on tools too", and the switch's second reading beside the selector (the valve's own switch arm): `ToolExecuteContext.toolSelector.admits` and the `layers` a child inherits used to answer without the switch while the valve answered with it. Added: none — the run's own layer is built in one place, where the switch is an allowlist naming nothing.
|
|
1485
|
+
- **Ceiling.** The layers a switched leg hands a child now carry its allowlist naming nothing (a child under them keeps only what it relies on). No engine lane spawns from a switched leg — the delegation and workflow tools are not mounted under it — so no child reads them today.
|
|
1486
|
+
- **Pins.** `test/backlog4356-unmount-keeps-relied-tools.test.ts` — U1 every protocol row × leg kind × ceiling × own list: the switched selector (built as prepare builds it) admits what an `allowTools: []` one admits, and the valve reads it · U2 a relied-on row stays; a teammate's row goes under a ceiling that denies it · U3 everything else goes, peers included, whatever the own allowlist named · U4 the ceiling a switched leg hands a child carries the switch · U5 prepare and the static plan build through the same reader · U6 / R7b a malformed allowlist beside the switch refuses as without it (also with a judgment of those lists, and in the static plan) · R12 the static plan under the switch · R13 no MCP dial under the switch · R1 / R2 the verifier (with and without an execution environment) is handed `StructuredOutput` alone, reads no file, its verdict lands · R3 control: without the switch the verifier keeps its read hands · R4 / R5 / R6b a structured run under the switch: `tools` = `["StructuredOutput"]`, the result carries the submission, a leg of no kind sends no `tools` key and reads `entries: []` · R6 control `allowTools: []` · R7 invalid schema refused · R8 own deny refused with and without the switch · R9 cascade passes on its first rung · R10 goal iteration keeps `DeclareDone` · R11 no delegation tool under the switch; a child under an empty allowlist keeps `ReportBlocked` alone · V1 a workflow agent under the switch keeps `ReportBlocked` (not with `enableBlockedReport: false`), delivers with a `{ schema }`, ends `blocked` on a call of it · V2 a leg the host starts as delegated: request and roster · V3 the governance baseline's `base` on the script path · V4 (current state, #4366) a headless all-ask policy reads a switched structured leg as it reads `allowTools: []` · V5 (current state, #4366) a parkable ask parks on the submission with its card · V6 a resume that re-gives the switch stays bare; one that leaves it out mounts the whole face (current state, #4367) · V7 (current state) a `verifierTools` entry the switch took leaves one plan row, absent `because: ["spec.selector"]` (the cell reads no notice sink). Re-pinned: `test/backlog2070-protocol-rows-under-ceiling.test.ts` P7 (a teammate row under the switch follows the ceiling — a reading that changed); mechanical, for the valve's one-argument signature and the switch built into the layer: `test/backlog1971a-valve-readers.test.ts`, `test/backlog1971b-mounted-set-and-valve-readers.test.ts`, `test/backlog1432-deny-universe.test.ts`, `test/backlog1433b-one-universe.test.ts`, `test/rescan-a087-fixes.test.ts`, `test/prepare-delegation-surface.test.ts`, `test/prepare-question-face.test.ts`, `test/backlog2001-refresh-teaches-once.test.ts`, `test/backlog3871-instructions-redial-digest.test.ts`, `test/backlog3717-roster-entry-bound.test.ts`.
|
|
1487
|
+
- **Consumers.** @server (`chat`): `excludeAllTools` + `verify` now judges and delivers a verdict; `excludeAllTools` + `outputSchema` delivers `structuredOutput`, so a refusal of that pair before sending can be lifted; a host that resumes a parked switched leg gives the switch again (Observable (d)). @server @sdk: a workflow agent or a governance baseline that carries the switch now hands its agents `ReportBlocked`, and such an agent can end `blocked`. @cli: an agent launched with an empty tool list (sent as `excludeAllTools`) delivers its structured result; its roster carries one row, `StructuredOutput`, when the request has a schema. @client-core: a roster under the switch may carry that one synthetic row. @test: "`excludeAllTools` ⇒ `entries: []` and the request carries no tools" holds for a top-level leg that relies on nothing; a structured leg carries `StructuredOutput`, a delegated leg `ReportBlocked`.
|
|
1488
|
+
- **Observable.** No key, no word, no frame kind moves. Under the switch: (a) the roster and the request carry `StructuredOutput` on a structured leg and `ReportBlocked` on a delegated one, and the static plan lists `StructuredOutput` mounted; (b) the event stream carries the submission's `tool_start` / `tool_end`; (c) the submission meets the host's ask policy, as it does under `allowTools: []`: a policy that asks for every tool parks the leg with a `StructuredOutput` approval card where a durable seat is wired, and without one refuses the submission — one more request, a `delegation.ask_unresolvable` notice, `completed` with no `structuredOutput` (before: one request, `completed`, no notice; #4366, KL-4356-d); (d) a resume reads the spec the host gives it again: one that leaves the switch out mounts the whole face (caller tools and hands) and completes (V5 / V6 assert the face and the outcome; no notice sink is read), as a resume that leaves out `allowTools: []` does — re-give the switch on every resume (#4367, KL-4356-e).
|
|
1489
|
+
|
|
1490
|
+
### A `resumeAt` landing (`at` / `before`) on a store that loads only a bounded window of the log is refused `resume_at.outside_window` when the store does not hold the whole branch it would continue — it no longer completes with the earlier conversation silently gone (#4359; P1, user-visible; narrowing: one, below, open to veto @server @cli @client-core @web-client; widening: none; hard-breaking: none; closed sets: `ConfigRefusalCode` +1; exports: none added, none removed; wire: one refusal word) (@server @cli @client-core @web-client @test)
|
|
1491
|
+
|
|
1492
|
+
- **Rule.** One judge for both words, after the target checks and before the leaf moves: the store must hold the whole branch the landing continues. It reads only how the store's own path read from the landing (the leaf `at` / `before` would set) ends — at a root ⇒ held; at the store's floor ⇒ held iff the latest compaction on the path keeps from an entry on it (the context of a branch starts at that kept entry either way, the entry `boundedTail` loads from); the store does not hold the landing (`not_found` from the path read) ⇒ not held. It never reads below the floor (a bounded load need not answer that). It fails closed and reads the store contract's words, not error classes: `corrupt_session` (the word the session store contract holds every store to for a parent missing above the floor and for a chain that loops) is left to the run, which meets the same damage and answers it as before; any other failure of the read — an I/O error, `conflict`, a word it does not know — is thrown as it came: the request fails on it before the leaf moves (it used to fail on it one read later, after a branch write and its write-back), and a lost race the acquire's retry recognizes (a `SessionError` with code `conflict`, the form the session store contract holds a race's loser to) goes back to that retry, which judges again; a value that only carries the word is not retried and fails the request. A read that failed once is never taken as "held". A store that answers a broken lineage with another word gets that word back as the request's failure (KL-4359-c).
|
|
1493
|
+
- **Before.** On a store whose wake loads the log from the newest compaction's kept entry on (the server's SQL session store), a landing between that floor and the compaction ran to `completed`, and that leg and every leg after it was sent the conversation from the floor on, with no summary. Measured on a store of that shape (real Runner, one compaction kept two turns, U-1 … U-6): `before` the user message between the floor and the compaction ⇒ the undo leg was sent 3 messages `[U-3, A-3, U-R]` (the whole log: 7, from U-1) and the next leg 5 (the whole log: 9); `at` it ⇒ 4 and 6 (8 and 10); `at` the floor message ⇒ 2 and 4 (6 and 8); two compactions, the same between the second floor and the second compaction ⇒ 3 and 5, 4 and 6 (8 and 10, 9 and 11, with a summary). A landing below the floor — `before` the floor message, `before` / `at` the opening row of a run that compacted inside itself, `at` a set-aside branch older than the newest compaction — was refused in the store's word: terminal `failed`, code `not_found`. The server re-measured both forms on its own SQL session store over real MySQL-protocol and PostgreSQL databases (cold and warm). On the documented bounded-tail recipe (`boundedTail` into a floored `InMemorySessionStorage`) the between-landing completed the same way (3 and 4 messages, no summary).
|
|
1494
|
+
- **Now.** Those landings are refused `resume_at.outside_window` at prepare: `TaskResult.terminal.kind: "failed"`, `terminal.code` that code, no model call, nothing written by the attempt that refuses, the leaf where it found it, and the next plain leg reads exactly as if no undo had been asked (pins A – C, G). A refused request that carries `restoreFiles: true` touches no file and makes no file-history call — the file restore runs after the acquire (the same request completed and restored the files before: the file moved back to its earlier content with one file-history restore call; pin J). One exception, a narrow write race: when an earlier attempt of the same request landed its leaf write, another writer then appended on the moved leaf and the retry was refused, the session stays on the branch that attempt opened, under the other writer's entries — the acquire's existing rule keeps what another writer built on rather than writing it back (measured with an injected race; `docs/INTEGRATION-CORE.md` §4.8). A landing the store holds — after the latest compaction, or on a branch whose own compaction defines the window — runs as before and is sent what the whole log sends (pins D, E).
|
|
1495
|
+
- **Narrowing (open to veto @server @cli @client-core @web-client).** On a bounded-window store, `at` / `before` whose landing's branch the store does not hold whole is refused — today that is a completed run that silently lost the earlier conversation (an Undo of a short run right after the next compaction kept its turns, a Redo onto such a point). Today's refusal in the store's word `not_found` for a landing below the floor is `resume_at.outside_window` now. Part of the same narrowing, on a log that is not damaged (@server @cli): a redo back onto the main branch after a compaction on a side branch that keeps from an entry below the main branch's own compaction — the newest compaction sets the floor and the main branch's kept entry is below it — read `corrupt_session` after a branch write and its write-back, and reads `resume_at.outside_window` with nothing written now (measured by the collection leg on a real compaction; pin Y). The set-aside-branch form of KL-4359-a ("a redo across a compaction on another line" whose branch forks below the floor) still reads `corrupt_session`: the two words coexist for one user action until the server's store picks its window by the landing's branch, which removes both. Part of the same narrowing, on a damaged log (@server): a landing whose path stops at the floor under a latest compaction whose kept entry the window does not show reads `resume_at.outside_window` with nothing written also when that kept entry is in no log at all — it was `corrupt_session` after a branch write and its write-back (pin H on the server's wake); the engine does not read below the floor, so "kept below the floor" and "kept entry missing" are one fact to it, and a plain leg on that branch still meets the damage (inferred, not measured: the context build refuses it, `src/engine/session/session.ts:233`). A whole-log store reads no change: its path always ends at a root (pins F, H, I, on the in-memory `TtlSessionStore`, the pins' one whole-log store: every landing completes, every target refusal keeps its code; that these readings equal the base tree's is the car's receipt reading, not a pin cell). These refusals go away by themselves once the server's store picks its window by the landing's branch at every `setLeafId` and answers its path read by the same rule (KL-4359-c (d)): such a store holds the branch and the landing reads as on the whole log (pins E, S) — except `at` a user message a cold handle does not hold, which #4376 (below) refuses with the same code until the store answers that message's group (KL-4376-a, #4377); a store that re-cuts only for a target outside its window keeps refusing the landings between the floor and the compaction on a cold handle (pin S).
|
|
1496
|
+
- **New code.** `resume_at.outside_window` (`CONFIG_REFUSAL_CODES`, `ConfigRefusalCode` 194 → 195). A host answers it like the other `resume_at.*` words — a 4xx with the code (the request moved nothing, except in the write race under "Now"), never degraded into a plain continuation or an `at` elsewhere.
|
|
1497
|
+
- **Not changed.** Every existing target refusal and its code (`resume_at.not_found`, `resume_at.not_a_message`, `resume_at.before_target_not_user`, `resume_at.before_root_unsupported`), judged first as before, below the floor too (pin I); a damaged log's answers on a whole-log store (a missing middle entry with no replay loss is `corrupt_session`, a replay-proven loss re-roots and is announced; pin H); a walk that breaks above the floor on a bounded store — still `corrupt_session`, including a set-aside branch that forks below the floor but was written after it (Redo onto it after a compaction on the main branch; the whole log completes it), because telling "below the floor" from "missing" would need a read below the floor; the store windowing by the landing's branch removes that form; the bounded-tail recipe's `before` on its floor message (`before_root_unsupported`, the parent is not loaded) and a target below its floor (`resume_at.not_found`); `TaskResult` members; the durable resume's rewind.
|
|
1498
|
+
- **Rules.** +1 judge (one function, one read, both words); no per-word and no per-store branch. "Held" is read from the fact the context is built from (the latest compaction's kept entry on the path), not from a second reading of the window.
|
|
1499
|
+
- **Known limits.** KL-4359-a: a redo onto an earlier branch after the main branch compacted, on a bounded-window store, reads `corrupt_session` (the whole log completes it); the session is not damaged — the refused request leaves two leaf rows off every branch, the leaf back on the main tip, the next leg reads as if nothing was asked, and the same log read whole redoes it (measured, pin H-window, a status-quo pin) — the store's fix (window by the landing's branch). KL-4359-b: a warm store handle whose window was whole when the leaf moved can accept a landing that a later cold wake of the same store does not hold — the store's fix, measured and in hand at the server. KL-4359-c: three things a store you ship keeps for the judge (`docs/INTEGRATION-CORE.md` §5.5) — a floor keeps its stored `parentId` (a floor rewritten with an empty parent reads as a root, and the between-landing loses the conversation as before), a leaf outside the window answers `not_found` (a one-entry path reads as a floor stop and the sentence calls the landing the floor), a missing or looping parent answers `corrupt_session` (`not_found` there is refused `outside_window` naming the wrong entry), and (KL-4359-c (d), @server) a bounded store answers its path read by the same window rule as its `setLeafId` — for a landing `setLeafId(x)` would hold, `getPathToRoot(x)` answers that path, read-only. A store whose `setLeafId` re-cuts its window by the target's branch while its path read answers `not_found` outside the window it holds now has every undo / redo it could hold refused `outside_window` on a cold handle and held on a warm handle that caches the whole log (pin S, status quo, not the goal: rows 1 – 6 of the refusal table, nine landings, nine refused cold with nothing written, nine held warm). The rule holds what `setLeafId` holds and no more: where `setLeafId` re-cuts only for a target outside its window, a landing between the floor and the compaction (the floor message included) stays refused cold and held warm, while one below the floor or on a set-aside branch is held warm, and cold but for `at` the opening row of a run that compacted inside itself (refused cold since #4376, below: a user message the cold handle does not hold, KL-4376-a); a store that picks the window by the landing's branch at every `setLeafId` holds all nine warm and eight cold (the same `at` refused cold since #4376), with the whole log's requests (pin S). The bundled stores keep all four; the server's SQL stores (today's and the next) keep (a) – (c) by the server's reading of its code, not run here. KL-4359-d (@server) — closed in the same release by #4376 (below): on a store that keeps (d), `at` a user message with an attachment frame below the floor completed and continued under the message, not under its frame (the inclusive mode looked for the frame among the entries the handle holds); it is refused `resume_at.outside_window` on a cold handle now, nothing written. The engine's fix (find the frame on the landing's branch) is a ticket of its own.
|
|
1500
|
+
- **Consumers.** @server: give `resume_at.outside_window` a row (4xx + code; the undo / redo endpoint says the step cannot be taken on this session as stored); your store ticket (the window by the landing's branch) makes these refusals disappear when the store's path read answers by the same window rule as its `setLeafId` and the window is picked at every `setLeafId`, not only for a target outside it (KL-4359-c (d), pin S) — all but `at` a user message a cold handle does not hold (#4376 below, KL-4376-a; lifted by #4377). @cli @client-core @web-client: on a long, compacted session behind the server, Undo / Redo of such a point is refused with this code instead of continuing without the earlier conversation — show it as a refusal, leave the transcript as it was. @test: black-box — a compacted session on the server, `before` the user message between the floor and the compaction ⇒ the refusal, and the next plain message's first model request carries the summary.
|
|
1501
|
+
- **Pins.** `test/backlog4359-resume-at-outside-window.test.ts` (real Runner, a scripted model, three stores: the whole log, the server's wake shape, the bounded-tail recipe; each cell read on the base tree and the head) — A one compaction, `before` / `at` between the floor and the compaction ⇒ refused, untouched, the next leg = no undo · B two compactions, the same between the second floor and the second compaction · C the floor message (`before` below it, `at` on it) and a run that compacted inside itself (`before` / `at` its opening row) · D after the latest compaction ⇒ held, the request the whole log's · E a store that picks its floor by the landing's branch ⇒ the set-aside branch held and read as on the whole log; a floor below the branch's kept entry held too; the global floor ⇒ refused · F the whole log under A – C: completed, unchanged · G the recipe: after its compaction held, between refused (never `corrupt_session`), below its floor `resume_at.not_found` · H a missing middle entry `corrupt_session`, a torn record re-rooted; on the server's wake a set-aside branch forked below the floor `corrupt_session` as before, a damaged latest compaction above a lower floor `outside_window` · I twelve existing target refusals on the whole log and on the server's wake, below the floor, unchanged · R a failed read of the landing's path (an injected one-shot I/O error on the server's wake and on the recipe, `at` / `before` between the floor and the compaction, and `at` a held landing) ⇒ the request fails with that error, no model call, nothing written, the leaf unmoved, the next leg = no undo; an injected `SessionError("conflict")` on that read ⇒ the acquire retries (one acquire more than the same request with no race) and the retry judges again (refused between, completed and sent what the whole log sends on a held landing; two tests); a parent chain that loops ⇒ `corrupt_session`, the same answer and rows as before · Y the redo back onto the main branch on an intact log ⇒ `outside_window`, nothing written (the whole log completes it) · J `restoreFiles: true` on a refused landing ⇒ the file stays, no file-history call · S the store read contract (KL-4359-c (d); a fixture store that floors a wake at the leaf's branch and re-cuts its window by the target's branch at `setLeafId`; a cold handle and a warm one that holds the whole log; rows 1 – 6 of the refusal table, nine landings, against the whole log): the path read by the rule agrees with `setLeafId` on every entry and moves no floor, the path read by the current window disagrees exactly on the seven entries below the floor; by the rule with a re-cut only outside the window ⇒ the five landings between the floor and the compaction refused cold and held warm, the four below the floor and on the set-aside branch held warm with the whole log's requests and cold but for `at` the run's opening row (refused cold since #4376: a user message the cold handle does not hold); with the window picked at every `setLeafId` ⇒ the same nine held warm, eight held cold; by the current window (status quo) ⇒ all nine refused cold, held warm; S frames (#4376): `at` a user message with an attachment frame below the floor ⇒ the whole log `[U-1, frame, U-R]`, both rule-keeping fixtures refused cold with nothing written (before #4376 they completed `[U-1, U-R]` with U-R under the message), the current-window fixture refused. Base tree: 5 of 7 red (the first seven tests), H and I green by design; the eighth (H on the server's wake) read with the judge off: the side branch the same, the damaged compaction `corrupt_session`; the five added tests read with the judge off: R's failed read ⇒ the same error after a branch write and its write-back, R's `conflict` ⇒ completed with the conversation from the floor on and no summary, R's loop the same as head, Y `corrupt_session` with two rows written, J completed with the file moved back. Mutations, each read per test: the judge removed ⇒ A, B, C, E, G red; "the kept entry on the path" widened to "the window holds a compaction" ⇒ A, B, G red; the floor stop classified by `getEntry` of the parent (unreadable ⇒ `corrupt_session`) ⇒ A, B, G red; narrowed to "the kept entry is the floor" ⇒ E red; the path read through the leg's replay-loss view ⇒ H red; the judge before the target checks ⇒ A, I red; `not_found` from the path read not judged ⇒ A, C, E red, and S's status-quo cell (its rows 4 – 6 held cold: that store's `setLeafId` holds them); S's own fixtures: its path read by the rule turned into the current window's ⇒ S's fixture check and its rows cell red (every cold landing refused); its window picked only outside instead of at every `setLeafId` ⇒ the rows cell red (rows 1 – 3 refused cold); the warm handle cut at a wake instead ⇒ the rows cell and the status-quo cell red (the warm landings read as the cold); every throw of the path read refused ⇒ H red; any other failure of the read taken as "held" (returned, not thrown) ⇒ R's failed read and both of R's `conflict` tests red (the held one alone: one acquire, as with no race); the lineage word read by error class (`instanceof LineageBreak`) instead of the contract word ⇒ R's loop red (0 rows where the run lays 1); `corrupt_session` thrown by the judge instead of left to the run ⇒ H (both tests) and R's loop red.
|
|
1502
|
+
|
|
1503
|
+
### An inclusive `resumeAt` (`at`) at a user message the session's handle does not hold is refused `resume_at.outside_window` — it no longer continues a branch without the message's attachment frames (#4376; P1, user-visible on a store that keeps KL-4359-c (d); narrowing: one, below, open to veto @server @cli @client-core @web-client — on today's stores no change; widening: none; hard-breaking: none; closed sets: none; exports: none added, none removed; wire: none — an existing code, one new refusal sentence) (@server @cli @client-core @web-client @test)
|
|
1504
|
+
|
|
1505
|
+
- **Rule.** `at` a user message keeps the message's group of attachment frames: the leaf is the group's last frame. The group is read by its one reader (`lastAttachmentFrameOf`, `src/core/user-attachment.ts`) among the entries the session's handle holds, and the reader now answers "cannot be read" when those entries do not hold the message (it used to answer "no frames"). The window judge (#4359, `src/core/runner/branch-window.ts`) judges the message's own branch first, then refuses an unread group with `resume_at.outside_window` before anything is written: no model call, no row, the leaf where it was, the next plain leg as if nothing was asked.
|
|
1506
|
+
- **Before.** On a store whose path read answers by the window its `setLeafId` would hold (KL-4359-c (d) — the form the server's next session store takes), `at` a user message whose group is below a cold handle's floor completed and continued under the message: the model was not sent the attachments (the whole log `[U-1, F-1a, F-1b, U-R]`, such a store `[U-1, U-R]`), and nothing said so. A store that floors at the group's last frame (the group across the floor) completed the same way (read from code: with the floor at an earlier frame of the group, the old reader found the frames above the floor and the store arm refused the landing).
|
|
1507
|
+
- **Now — the refusal sentence.** `resumeAt entry "…" keeps a user message whose attachment frames cannot be read: the entries session "…"'s handle loaded do not include that message (the store answered it by id), so where its attachment frames end is not known and the branch would continue without them — retry on a handle that holds the message (the whole log, or the message's own branch), or rewind with resumeAtMode "before" to drop the message and send it again with its attachments`. The store arm's sentences are unchanged.
|
|
1508
|
+
- **Narrowing (open to veto @server @cli @client-core @web-client).** On a store that keeps (d), `at` a user message a cold handle does not hold is refused with or without frames: the handle cannot tell "none" from "not loaded" — the message's row carries no count of its frames (read: `UserMessage`, `src/engine/llm/types.ts`; the frames are minted as their own rows, `src/engine/harness/agent-harness.ts`), and a group can be the instruction files a leg's first load delivered (`src/core/runner/run-leg.ts`, the trailing frames). A warm handle that holds the message is unchanged. These are fixture readings (stores that re-cut for a target outside the window, for every target, over the branch only — the refusal does not depend on the trigger: the group is read before the leaf moves); the server's deployment is not read here: its own store's readings are the server's to take. **On today's store shapes nothing changes** (measured on fixtures): the bundled whole-log stores hold every message; a fixture store shaped like the server's current SQL wake, and every store whose path read answers `not_found` outside its window, already refuse these landings in the judge's store arm — same code, the store arm's sentence, nothing written (read from code: for a group across a floor at its first frame the store arm's reason reads "does not hold entry" where it read "stops at entry" — the landing is now the message, not its last frame).
|
|
1509
|
+
- **Not changed.** `before` (the leaf is the message's parent: the group leaves with it) and `at` an assistant message (no group is read) — measured on every store, cold and warm; `at` a custom message reads no group either (read from code: only a user message's group is read, `prepare-acquire-reconcile.ts`), not measured; `at` a user message the handle holds (its group inside the window) — the whole log's request.
|
|
1510
|
+
- **Rules.** The reader: still one, an answer ("cannot be read") in place of an implicit "none"; the judge: +1 verdict (an unread group is not held). No new read, no second lookup, no new write, no new code, no new store requirement.
|
|
1511
|
+
- **Known limits.** KL-4376-a: the narrowing above — lifted by an optional store read that answers an entry's children by the window its `setLeafId` would hold (#4377). KL-4359-d closed.
|
|
1512
|
+
- **Consumers.** @server: no store change asked; on your next session store (path read by (d)), `at` a user message below a cold handle's floor answers this refusal instead of a run without its attachments, and one cell of #4359's rows 4 – 6 cold reading changes (`at` the opening row of a run that compacted inside itself: refused cold, held warm). @cli @client-core @web-client: the existing `resume_at.outside_window` row covers it — show the refusal, leave the transcript as it was. @test: black-box — a session whose first message carried attachments, compacted behind a store that keeps (d), `at` that message on a fresh wake ⇒ refused, transcript unchanged; on a whole-log session ⇒ the request carries the attachments.
|
|
1513
|
+
- **Pins.** `test/backlog4376-resume-at-attachment-group-outside-window.test.ts` (a real Runner, a scripted model; stores: the whole log, the server's wake today, the bounded-tail recipe, two that keep (d) — re-cut for a target outside the window / for every target —, one that keeps (d) over the branch only, one whose path read answers its current window; cold and warm handles) — `at U-1` (two frames, below the floor) · `before U-2` · `at A-1` · `at U-2` (no frames) · `at U-5` (a group above the floor) · the group across the floor (today's store, and one that keeps (d): before, floored at the last frame, it completed without either frame); `test/backlog4359-resume-at-outside-window.test.ts` S rows and S frames re-pinned. Mutations: the reader's "cannot be read" turned back into "none" ⇒ the three cold `at U-1` cells complete without the frames; the refusal moved after the leaf write ⇒ every group-arm cell writes two rows; the group arm given the store arm's sentence ⇒ red.
|
|
1514
|
+
|
|
1515
|
+
no-notes: #4110(readings only: a test file measures which store waits between a park and the published result have no time bound; nothing in the package changes), #3098(pins only, zero src: `test/backlog3098-interrupted-edit-restore.e2e.test.ts` holds that an Edit whose write is cut off — a real SIGKILL inside the write on a first and a later turn, or an interrupt that lands mid-write — leaves the turn's boundary committed before the first byte and the first-touch record kept, and that a new process restores the turn-start bytes by the prompt's entry id through `rewindFilesTo` and `Runner.rewindFiles`; the 404 a shell sees after a kill is answered before the engine is reached, by the deployment's run-id-to-entry translation; one flaky-guard allowlist entry for the fixture's widening delay), #3008(named by a #1530 commit subject only: the #3008 pins' "as made" reading looks past the admission seat's own `origin` stamp; test file only, the stamp is told in the #1530 section), #562(named only by a car-small-7680-c commit that re-pins test/backlog562-manifest-mcp-face.test.ts to read the unauthorized row's #3984 clause; nothing for #562 changes), #4079(named only by the car-small-7680-c commit that re-reads the backlog3981 X14 / X15 controls by the #4079 reading shipped in 7.67.0; nothing for #4079 changes here), #1440(named only by a batch-1 bookkeeping commit that points its contract cell at the 7.67.0 notes; nothing for #1440 changes), #1614(the same commit and the same reason as #1440), #3010(named only by the batch-1 commit that re-reads the mods-s4 E1 equivalence shape after #1530; nothing for #3010 changes), #3011(the same commit and the same reason as #3010), #3943(named only by the batch-1 commit that re-reads the fork-lane cell of its pin by #3989; nothing for #3943 changes), #4054(named only by the batch-1 commit that re-pins its refusal fragment to #4012's sentence; nothing for #4054 changes), #4111(named only as the ticket that closes the transitional fork-leg cause read by the batch-1 #4079 re-pin; nothing for #4111 lands here), #4137(release tooling only, scripts/release/ and its declaration twins, nothing in the package: the release re-window reads a RESOLVED-PENDING-MERGE row whose own car's head did not decide by the commits its status head names, so a row whose commit is in the release stays at the cut window and a row whose commit an earlier release carried stays with the flip warning naming that release; the rows read by the car's name alone and the rows kept on a commit are listed one line per row in the dry run; the membership readers moved to scripts/release/membership.mjs and lib.mjs re-exports them), #4140(repository tooling only, scripts/rulings-gate.mjs and scripts/rulings-disposition.mjs: the rulings lookups call a RESOLVED-PENDING-MERGE ticket landed only when the commits its row names are in the tree being judged, by the same reader the re-window uses, instead of by the status word alone), #4136(a background child's launch card, its listed row and its approval already carry one registry handle; the change is a sentence in INTEGRATION-CORE §1.6a saying siblings are told apart by `agentId`, not by the model-chosen `sourceAgentName`, and its pin — no behaviour), #1698(pins only: the wake coverage table per task type on the real Runner — every type the engine starts parks its completion after an idle leg and a `wake` shows it once; the second block of the ticket's reading is a host's own copy written into the wake turn's objective, KL-1698-a; no src), #4104(measurement cells only: `test/backlog4104-parked-leg-stuck-child.test.ts` reads what bounds a parked leg's result while a covered background agent sits in a call that ignores the abort — no src, no behaviour; the fix is its own ticket's notes), #3826(the author table's fixture gains one row for the new `inlineUntrusted(words)` call site of the SendMessage receipt label; a test fixture, no src, no behaviour), #4170(a source comment only: the comment on the `Monitor` tool's event sources says the `ws` arm is this engine's choice, not an alignment; zero behaviour, nothing in the package changes), #4197(docs and a pin only, zero src: INTEGRATION-CORE 10.6a says which seat reaches a delegated child's SubagentStart / SubagentStop — `RunnerDeps.mods` follows the Runner the child's leg runs on, `TaskSpec.mods` follows the spec — with the lane readings, and `test/backlog4197-subagent-events-lanes.test.ts` pins them on real Runners; nothing in the package changes), #4184(the TaskResult field-set pin registers `limitCrossed` with its review note; a test file, no behaviour), #1367(the inline-untrusted author table gains the two notice echoes of `auto-strip-notice.ts`, sink log, author e; the auto-strip source vocabulary is spelled through one named alias `PersonalRuleSource`, `RuleSource` = `org` | it, and the person-path rows' local twin folds into it, so the release manifest reads the set; the published faces keep their declared `Exclude<RuleSource, "org">`, a derivation of the same words; no behaviour, no wire), #4170(the src/tools line-anchor quota 16 -> 17 for the reference corpus coordinate its comment carries, with the reason; a test file), #4209(the nested lane's thrown-open cells re-read through the #4040 `openBoundAt`, which answers a throw as the refusal `unknown`: the member is not read and the operator's unreadable-file sentence is said; test file only), #4167(pins only, zero src: `test/backlog4167-write-unreadable-existing.test.ts` holds that a Write onto an existing file the engine cannot read is a tool error the model answers in the same task, with nothing written — the reading the reference has; two secondary differences are known limits — KL-4167-a, the card before the error, registered as D-4167-i, and KL-4167-b, the read-first sentence, registered as D-4167-ii), #1643(named in the subject of a #4157 commit that holds the module-size ceiling of shell-run-candidates; nothing for #1643 lands here), #3757(test-only: the raw-splice census drops two log rows whose sites the notice throat took over; the notes of that work are under #3832), #1314, #1364, #734(named only by #4111's commits as the pins it re-pinned — the way-out remedy, the observer under a mandate, the sixth `parkLaneArmed` site; their behaviour is the #4111 entry above), #3613(named in the subject of a #4111 bookkeeping commit: the stage's pin of the CheckpointErrorReason count moves 36 -> 37 with leg_not_resumable; the notes of that member are under #4111), #4216(a test-only commit after the release pins that the catalog door's refusals are exactly four codes; the notes of the door shipped in 7.67.1), #3202(named in the subject of a #3970 commit that banks the unknown-key gate ratchet 108 -> 107 once the notation key check retired; nothing else of #3202 lands here), #3355(named only by a #4234 commit subject for the two of its pins that were re-pinned — S16 and the child gate; the change is #4234's and its notes are there), #1050(named only in the subject of a #4280 fix2 commit, d10555087, that re-pins `test/backlog1050-file-history-capture-off.test.ts` to #4280's reading of the history root — read only when a point is recorded or restored; no #1050 code or behaviour changes here), #1623(named only in the subject of a #4357 commit, 3cdebcc32, that re-pins one cell of `test/backlog1623-tool-selector.test.ts` to #4357's reading — a deny list naming the whole peer leaves nothing to dial, so no `a2a:` warning on that side; no #1623 change lands here), #4131(release-gate tooling and pins only, zero src and nothing in the package: `scripts/run-live-gate.mjs` reads a written live file list `test/live-gate-files.json` that must equal the disk both ways, the behaviour cells read by majority, a request tap for the skills and P7b cells; `docs/TESTING.md`)
|
|
1516
|
+
|
|
1
1517
|
## 7.67.3 — 2026-11-12
|
|
2
1518
|
|
|
3
1519
|
### `providerMessage` is no weaker than the same response's `terminal.message`: what the raw-body reading withholds inside the sentence's literal, the sentence withholds too — a token split inside by a line break, a tab or an escaped Unicode space no longer rides the retry frame or the terminal refusal account in part (#4343; Security; behavior change: narrowing — more of such a sentence renders `[redacted]`, see Legibility; widening: a value read inside an engine tag now leaves the rest of the sentence readable, see Now; hard-breaking: none; closed sets: none; exports: none from the package root; wire: none) (@server @cli @client-core)
|